Knowledge spillovers from clean and dirty technologies · 2017. 10. 13. · Knowledge spillovers from clean and dirty technologies⇤ Antoine Dechezleprêtre† Ralf Martin‡ Myra

Knowledge spillovers from clean and dirty technologies

⇤

Antoine Dechezleprêtre† Ralf Martin‡ Myra Mohnen§

12th October 2017

Abstract

Government policy in support of innovation often varies across technology areas. Animportant example are climate change policies that typically try to support so calledclean technologies that avoid greenhouse gas pollution and hamper dirty technologiesthat are associated with polluting emissions. This paper explores the economic conse-quences of such policy moves in the short run. At the margin private returns of R&Dinvestments in different areas should be equalised. Hence, shifting the composition ofR&D activities by a policy intervention will only have a meaningful impact on economicoutcomes if the external returns differ. Hence, we compare innovation spillovers betweenclean, dirty and other emerging technologies using patent citation data. We developnew methodology including the usage of Page rank measures developed by Google torank web content. Exploring a wide range of robustness checks we consistently find upto 40% higher levels of spillovers from clean technologies. We also use firm-level finan-cial data to investigate the impact of knowledge spillovers on firms’ market value andfind that marginal economic value of spillovers from clean technologies is also greater.

⇤We thank Baran Doda, Jason Eis, Carolyn Fischer, Timo Goeschl, David Hemous, Colin McCormick,David Popp, Sjak Smulders and Nikolas Wölfing for many helpful comments. Participants at seminars at UKEnvEcon, AERE, EEA, WCERE, LSE, Harvard, Dublin, Geneva, Oslo, Brussels, Berlin, Bern, Mannheim,Zurich and SPRU have all improved the paper. The research leading to these results was funded by the SwissNational Science Foundation under the Sinergia programme, Project No CRSII1 147612. Other financialsupport has come from the Grantham Foundation for the Protection of the Environment, as well as the UKEconomic and Social Research Council through the Centre for Climate Change Economics and Policy.

†Grantham Research Institute on Climate Change and the Environment, London School of Economics,[email protected]

‡Imperial College and London School of Economics, [email protected]§University College London, [email protected]

1

1 Introduction

It is commonly recognized that knowledge spillovers from innovative activities provide a case

for government intervention in the market because private R&D investments are likely too

low. It has also been recognised that not all innovations create spillovers to the same extent.

In particular more basic research is assumed to create stronger spillovers and therefore should

attract more government support. However, for better or worse governments often champion

specific technology areas – rather than types - such as defence, IT, aerospace, bio technology

etc. Often this is because a certain area promises auxiliary – i.e. not necessarily economic –

benefits such as security, health or simply prestige. If the degree of spillovers that is generated

by these different areas are the same, then – from an economic point view – we don’t need to

worry about this. However, if there spillovers vary substantially across areas the distribution

of government intervention can affect the level and growth of economic well being. To the best

of our knowledge this study is the first to systematically compare spillovers between different

technology areas. Our main focus is what we have dubbed dirty and clean technologies; i.e.

technologies that are associated with GHG gas pollution and alternative technologies that

can replace them. However, we also examine other emerging technologies and we develop

methodology that will be relevant to comparing spillovers between technology areas more

widely. We focus on clean and dirty technologies because they are an important example

of deliberate differential treatment of technology areas by government policies. Increasingly,

governments are deployment carbon pricing policies which incentivise clean and hamper

dirty technology development. This includes carbon and energy taxes (Aghion et al 2016)

but also direct subsidies for clean innovation. In 2012, OECD countries spent over 3 billion

euros to support the development of new clean technologies such as renewable energy or

hydrogen cars (reference?). The motivation for this is the desire to mitigate climate change

in the long run. However, many policy makers – often in an effort to make climate change

2

policy attractive to the public - have suggested that this could also have a beneficial impact

on economic outcomes such as growth or employment in the short run. Theoretically, this

can only be the case if clean technology innovation leads to larger spillovers than the dirty

technology innovation that it replaces. Hence, the main objective of this paper is to measure

and compare the amount of knowledge of spillovers from clean and dirty technologies.

Following a long tradition in the literature, we derive our measure of knowledge spillovers

from patent citation data (Trajtenberg (1990); Caballero and Jaffe (1993); Jaffe and Tra-

jtenberg (1999); Hall et al. (2005)) although we go beyond the usually used citation count

measures.

Patent documents offer a paper trail of knowledge flows as inventors are required to reference

previous patents which have been useful for developing the new knowledge described in the

patent. Patent citations are not without limitations, but an important advantage of our

dataset is that it allows us to deal with most of the problems usually associated with their

use. For example, we can identify (and discard) self-citations by inventors, as well as citations

added by patent examiners, which might not capture external knowledge spillovers. We rely

PATSTAT a new dataset assembled by the European Patent Office in collaboration with the

OECD. It provides information on nearly all patents filed worldwide in almost all national

patent offices. It also provides information on patent families; i.e. when the same innovation

is filed repeatedly in different jurisdictions. This allows us to use an innovation, rather than

a patent as the unit of analysis avoiding any double counting. Our main analysis focuses on

two technology fields - cars and energy generation - and within each field on two main areas:

fossil fuel based technologies (dirty) and alternative (clean) technologies. Cars and power

generation account for about 40% of global carbon emissions (IPCC, 2007) They also allow

an easy distinction between dirty - i.e. everything related to fossil fuel combustion - and

clean - i.e. alternative technologies such as electric vehicles and solar power generation. As

an extension we also consider “grey” technologies; i.e. innovation to improve the pollution

3

efficiency of fossil technologies, although this is harder to classify.

There are a variety of confounding factors that might lead to differences in citatations be-

tween technology areas such as clean and dirty that are un-related to spillovers in an economic

sense. Citations in patent documents are driven by legally binding definitions on what con-

stitutes prior art. These differ over time and between jurisdictions. Citation behaviour might

also vary over time because of technological developments such as the

Clean and dirty innovations are not uniformly distributed neither across space nor time.

Hence, average citation counts for the two technology areas might differ for example because

one areas tends to be filed more with patent offices that require more citations. We address

a wide range of such concerns by including a wide set of control variables such as patent

office by year fixed effects.

However, this will not deal with variation in citation practice between different technological

areas; e.g. suppose that in some technological areas it is customary to cite more frequently

by explicity referring to more remote underlying ideas. Because most innovations receive

their citations from within their own technological area, this could lead to differences in

citation numbers that reflect “cultural” differences between technological fields rather than

than economically meaningful spillovers. We address this in two ways. Firstly, we exam-

ine spillovers differences relying only on citations outside an innovation’s technoligcal area.

Secondly, we the Page-rank measures of an innovation rather than the mere citation count.

The Page-rank was developped by Google’s Larry Page to rank the relevance of webpages

on the basis of how they are hyperlinked; i.e. cited. We are one of the first to apply this

to patent data. It is recursively computed as the weighted average of all citing patent page

ranks weighted by the inverse ratio of citations in a citing patent.1 Hence, a patent receives

a high page rank if it is cited by many other patents that are themselves cited a lot but do

not cite many others themselves. This not only deals with potential variation in citation cul-1We discuss this in more detail below.

4

ture between technological areas but also considers indirect spillovers; i.e. an innovation can

create spillovers because it is cited a lot by itself or because it is cited by another innvation

that is cited a lot.

Our results suggests that clean innovations generate singificantly more knowledge spillovers

than their dirty counterparts. All other things being equal, clean patented inventions receive

43% more citations than dirty inventions. The gap is larger in the electricity production

sector (49%) than in the transportation sector (35%). Interestingly, the gap between clean

and dirty technologies has been constantly increasing during the past 50 years. We show

that clean patents are not only cited more often, they are also cited by patents that are

themselves cited more often (irrespective of their technological area). When considering

our new PatentRank index, we also find strong evidence of larger spillovers from clean

technologies. Our conclusions are robust to a large number of sensitivity tests. These

include discarding citations added by patent examiners, correcting for self-citations at the

applicant level, including inventor fixed effects, looking at different subsamples and including

additional control variables.

How can we account for the larger knowledge spillovers from clean technologies? One ex-

planation stands out from our investigation: clean technologies seem to benefit from steep

learning curves associated with new technological fields. 2When we control for the age of

the technology, the clean premium decreases by 14%. We then compare knowledge spillovers

between clean, grey and “truly dirty” innovations. The analysis suggests a clear ranking:

clean technologies exhibit significantly higher levels of spillovers than grey technologies,

which themselves outperform truly dirty technologies. We also compare clean inventions

with other emerging technologies such as biotechs, IT, nanotechnology, robot and 3D, and

find that clean patents appear much closer in terms of knowledge spillovers to these radically2We partially control for this by including a measure of previous patenting within the technology class

of a given patent in our regressions, but this novelty effect might not be well captured by the number ofpatents.

5

new fields than to the dirty technologies they replace. Interestingly knowledge spillovers from

clean technologies appear comparable in scope to those in the IT sector, which has been the

driver behind the third industrial revolution. When comparing clean, dirty and emerging

technologies to all other inventions patented in the economy, we find a clear ranking in terms

of knowledge spillovers: dirty technologies have lower knowledge spillovers than the average

invention, while clean and other emerging technologies exhibit larger knowledge spillovers.

With the exception of biotechs, all other emerging technologies (IT, nanotechnology, robots

and 3D) show larger knowledge spillovers over the average invention than clean inventions.

Taken together, these pieces of evidence suggest that the clean advantage might be a feature

of the radical novelty of the field.

We make every effort to control for confounding factors in citation behaviour between clean

and dirty technologies. However, with patent and citation data alone we cannot infer any-

thing about the actual economic value of innovations and their spillovers. This could be a

problem if - say - dirty spillovers generated vastly more economic value per innovation than

clean spillovers which could imply that the economic relevance of dirty spillovers is higher

despite their lower frequencey as measured by citations. For a subset of our data we can

address by expanding an approach first introduced by Hall et al. (2005); i.e. we look at the

stock market values (Tobin’s Q) of firms taking out patents. Specifically, we examine if the

value of a new innovation on a firm is higher if the innovation is benefitting from a dirty

as opossed to a clean spillovers. However, our results suggest the opposite: a firm’s value

increase from a new innovation is higher when benefiting from more clean spillovers. Hence,

this re-inforces the spillover advantage of clean technologies.

Our results have a number of immediate implications. Firstly, they highlight that there

large and economically relevant spillover differences between technology areas and therefore

a meaningfully growth policy design should take these into account.

Secondly, with respect to climate change policy, our findings provide support for the idea that

6

pollution pricing should be complemented with specific support for clean innovation—e.g.

through additional R&D subsidies—that goes beyond standard policies in place to inter-

nalize knowledge externalities. Indeed, the higher spillover effects from clean innovation

compared to dirty innovations (including “grey” energy efficiency technologies) uncovered in

this paper justify higher subsidies to clean R&D in a first best policy setting. Radically new

clean technologies should receive higher public support than research activities targeted at

improving on the existing dirty technologies. However, such specific support could equally

be justified for a range of other emerging areas, such as nanotechnologies or IT. Therefore

our results go some way into supporting the recommendation by Acemoglu et al. (2012) that

only clean (and not dirty) technologies should receive R&D subsidies.3

Thirdly, our results lend support to the idea that a redirection of innovation from dirty

to clean technologies reduces the net cost of environmental policies and can lead to higher

economic growth in the short run, if the benefits from higher spillovers exceed these costs.

Indeed, if the factors leading to an under-provision of knowledge are more severe for clean

technologies and if new clean technologies are induced by environmental regulation, environ-

mental policies could generate growth by unintendedly correcting a market failure that has

been hampering the economy, irrespective of the environmental problem (Neuhoff (2005)). In

fact, the presence of a market failure associated with R&D spillovers from clean innovations

is one of the possible theoretical foundations for the Porter hypothesis (Porter and Van der

Linde (1995)) according to which environmental regulations may enhance firms’ profits and

competitiveness (see Ambec et al. (2013) and Ambec and Barla (2006), for a recent review).

For example, in Mohr (2002), the existence of knowledge spillovers prevents the replace-

ment of an old polluting technology by a new, cleaner and more productive technology, as3Interestingly, though, for a reason that is not present in their model: Acemoglu et al. (2012) do not

assume different spillovers from clean and dirty technologies. The crucial assumption on which the resultsby Acemoglu et al. (2012) hold is that patents last only for one period. Greaker and Heggedal (2012) showthat it is possible to obtain similar results when relaxing this assumption if one now assumes that cleantechnologies exhibit larger knowledge spillovers than dirty technologies.

7

firms have a second-mover advantage if they wait for someone else to adopt. The introduc-

tion of an environmental regulation induces firms to switch to the new, cleaner technology.

This simultaneously improves environmental quality and eventually increases productivity.

Our results however suggest that the potential growth effects of environmental policies very

much depend on the type of displacement being induced by increasing support for clean

technologies. If this leads to less investment in dirty technologies, as evidenced by Aghion

et al. (2012), there seems to be scope for medium run growth effects. If innovation in other

emerging areas is crowded out, such effects are less likely.

Our results also have implications for the modeling of climate change policy. For example,

Fischer and Newell (2008); Fischer et al. (2013) assess different policies for reducing carbon

dioxide emissions and promoting innovation and diffusion of renewable energy, with an ap-

plication to the electricity sector. They model R&D investments and learning-by-doing, but

assume that knowledge spillovers have the same intensity across clean and dirty technologies.

Our paper suggests that this assumption does not hold in practice and provides estimated

parameters that can be used to more precisely model the difference between clean and dirty

technologies.

Our paper relates to three main strands of the literature. First, our work draws on the

extensive empirical literature that has used patent data to analyze the determinants and the

effects of knowledge spillovers. Pioneers of patent citation data as a measure of knowledge

spillovers include Scherer (1965) and Schmookler (1966). Griliches et al. (1991); Griliches

(1992) survey this earlier literature. Since then, a large number of papers have used this

method to investigate knowledge diffusion (see, among others, Trajtenberg (1990); Caballero

and Jaffe (1993); Hall et al. (2001)). In particular, many papers have focused on the ge-

ography of knowledge spillovers (Jaffe et al. (1993); Jaffe and Trajtenberg (1996, 1999);

Thompson and Fox-Kean (2005)).

Second, in the energy literature some papers have recently attempted to compare knowl-

8

edge spillovers from energy technologies with those of non-energy technologies. Bjørner and

Mackenhauer (2013) compare the spillover effects of private energy research with those of

other (non-energy) private research. They find that spillover effects of energy research may

be lower than for other types of private research. Popp and Newell (2012) use US patent ci-

tation data to compare the social value of alternative energy patents to that of other patents

filed by the same firms. They find that alternative energy patents are cited more frequently

by subsequent patents, and by a wider range of technologies, than other patents filed by the

same firms. However, none of these papers distinguishes between clean and dirty technologies

within energy technologies.

Third, our paper is closely related to the literature on the impact of environmental policies

on economic growth, which is itself rooted in the endogenous growth literature (for seminal

contributions, see Romer (1990); Aghion and Howitt (1992, 1996, 1998); Grossman and

Helpman (1991)). Smulders and De Nooij (2003) introduce a difference in spillovers from the

clean and the dirty sector into a model in which both the rate and direction of technological

change are endogenous. They discuss the implication of this difference for growth in the long

run. In a Schumpeterian growth model where new technologies are both more productive and

more environmentally-friendly, Hart (2004) shows that environmental policy can stimulate

economic growth (see also Hart (2007); Ricci (2007b), for similar types of models, and Ricci

(2007a), for a review of this literature).

The remainder of the paper is organized as follows. In the next section we present the

datasets, explain how we measure knowledge spillovers and conduct some preliminary data

exploration. In section 3, we discuss our empirical strategy in greater detail. Section 4

reports our main results. In section 5, we estimate the market value of clean knowledge

spillovers. We discuss the implications of our findings in the final section.

9

2 Data and descriptive statistics

2.1 The patent database

We use data from the World Patent Statistical Database (PATSTAT), maintained by the

European Patent Office (EPO). PATSTAT includes close to 70 million patent documents

from 107 patent offices. We identify clean and dirty patents using the International Patent

Classification (IPC) and the European Patent Classification (ECLA). For this purpose we

rely heavily on work carried out at the OECD and the EPO, which has recently devel-

oped a patent classification scheme for "Technologies related to climate change mitigation

and adaptation" (see Veefkind et al. (2012) for more information on how this scheme was

constructed).4

We focus on two sectors where we can precisely distinguish between clean and dirty patents:

electricity production (renewables vs. fossil fuel energy generation) and automotive (elec-

tric and hydrogen cars vs. internal combustion engines). Our paper rests primarily on

a distinction between radically clean innovations (electric cars, solar energy...) and their

dirty counterparts (gasoline-fueled cars, coal-based electricity generation...). However, an

important feature of the dirty category is that some patents included in this group aim at

improving the efficiency of dirty technologies (for example motor vehicle fuel efficiency tech-

nologies), making the dirty technology less dirty. We refer to these energy-efficiency patents

as “Grey” inventions. The list of patent classification codes used to identify clean, dirty and

grey inventions is shown in table 20 and 21.

Given that the same invention may be patented in several countries, our level of observation

is the patent family (the set of patents covering the same invention in several countries). In4This new scheme was defined with the help of experts in the field, both from within and outside the

EPO, including from the Intergovernmental Panel on Climate Change (IPCC). It brings together technologiesrelated to climate change that are scattered across many IPC sections and includes around 1,000 classificationentries and nearly 1,500,000 patent documents.

10

Table 1: Number of clean and dirty inventions by sector

Sector Clean Grey True Dirty Total

Transport 74,877 133,083 212,193 420,153Electricity 103,659 19,827 627,590 751,076

Total 178,536 152,910 839,783 1,171,229

other words, we treat multiple filings of an invention as one invention and count citations by

patent family instead of individual patents.5 In total, our sample spans from 1950 to 20056

and includes over 1 million inventions with approximately 3 million citations made to these

inventions. A breakdown of the number of inventions in each sector can be found in table 1.

Clean inventions represent around 15% of our sample.

2.2 Citation counts as knowledge spillovers

Patent data has a number of attractive features. First, patents are available at a highly

technologically disaggregated level. This allows us to distinguish between clean and dirty

innovations in several sectors, in particular electricity production and transportation. In

comparison, R&D expenditures of a car company cannot usually be broken down into clean

and dirty innovations. Second, patent documents contain citations to "prior art" as inventors

are required to reference previous patents that have been used to develop the new technology

described in the patent. Citations are a response to the legal requirement to determine the

scope of an inventor’s claim to novelty and thus represent a link to the pre-existing knowledge

upon which the invention is built. 7In other words, a citation indicates that the knowledge5A patent family is considered clean if at least one patent within the family is clean6We stop in 2005 to allow at least five years for patent to get cited. The majority of citations occur

during the first five years of a patent.7US patent law 37 C.F.R 156 establishes that ’each individual associated with the filing and prosecution

of a patent application has a duty of candour and good faith in dealing with the (US Patent) Office,which includes a duty to disclose to the Office all information known to that individual to be material topatentability [...] no patent will be granted on an application in connection with which fraud on the Office waspracticed or attempted or the duty of disclosure was violated through bad faith or intentional misconduct’.

11

contained in the cited document has been useful in the development of the new knowledge

laid out in the citing patent and thus represents a knowledge flow (Collins and Wyatt (1988)).

It is therefore not surprising that patent data have been widely used in empirical studies

of knowledge spillovers (Jaffe et al. (1993); Jaffe and Trajtenberg (1999); Keller (2004);

Caballero and Jaffe (1993); Jaffe and Trajtenberg (1996)).

To give a concrete example of knowledge spillovers, take the patent entitled “X’Ray Appara-

tus” (US8036340B2, see figure 8). It was applied for in 2008, published in 2011 and belongs

to the H05K class of electric techniques. The patent documents the inventor(s), and the

applicant of the invention as well as their addresses. It also lists the claims of the invention

and references other patents which will be useful in the making of the invention, including

whether these citations were added by the examiner or not. Among its references, it lists a

patent US6727670B1 entitled “Battery Current Limiter for a High Voltage Battery Pack in a

Hybrid Electric Vehicle Powertrain” (see figure 7) which was published in 2004. It belongs to

the “electric motor” class (H02P). The citation received represents a transfer of knowledge.

Looking in turn at the list of reference, it cites the patent US6026921A (“Hybrid Vehicule

Employing Parallel Hybrid System, using both Internal Combution Engine and Electric Mo-

tor for Propulsion”, see figure 6) which was published in 2000 is classified as B60K which

falls under our clean transport category. This represents a clean knowledge spillover.

For each patent family in our dataset, we compile all the citations received regardless of their

In contrast, the EPO has no requirement similar to the duty of candour. Rule 42 of the European PatentConvention requires that the description in a European patent application should ’indicate the backgroundart which, as far as is known to the applicant, can be regarded as useful to understand the invention,draw up the European search report and examine the European patent application, and, preferably, citethe documents reflecting such art’. The different legal requirements of the two systems have implicationsboth in terms of who adds the citations and in the number of citations in the patents. For EPO patents,it is the patent office’s examiner rather than the inventors or applicants who adds the majority of patentcitations. This implies that in the EPO system, inventors are more likely to be unaware of the patents thatare (ultimately) cited in their patents. However, citations in EPO patents may be less ’noisy’ than USPTOcitations, since it can be assumed that they have been scrutinised and chosen by the patent examiner, andciting-cited patent pairs might be ’closer’ both in time and technological content than those extracted fromthe USPTO [Breschi and Lissoni (2005); Michel and Bettels (2001)]

12

field and whether or not they are clean.

Nevertheless, there are a few drawbacks to bear in mind. Patent citations are an incomplete

measure of knowledge flows because they only capture flows that result in a novel and

patentable technology. For this reason Griliches (1992) refers to citations as “pure knowledge

spillovers”. Since not all inventions are patented, patent citations underestimate the actual

extent of knowledge spillovers. Other channels of knowledge transfers, such as non-codified

knowledge and embodied know-how (inter-firm transfer of knowledge embodied in skilled

labor, knowledge flows between customers and suppliers, knowledge exchange at conferences

and trade fairs, etc.) are not captured by patent citations. It is however reasonable to assume

that knowledge spillovers within and outside the patent system are correlated. Furthermore,

there is a consensus that patent citations are a noisy measure of knowledge flows (Jaffe et al.

(2000)). First, citations made to patents by the same inventor (refered to as self-citations)

represent transfers of knowledge that are mostly internalized, whereas citations to patents

by other inventors are closer to the true notion of diffused spillovers. However, this problem

can be (at least partly) resolved by excluding self-citations by the inventor. Second, some

citations are added by patent examiners during the examination process (see Cockburn et al.

(2003) for an overview of the process). In a survey of inventors, Jaffe et al. (2000) show that

the influence of examiners on citations is considerable, and that inventors were fully aware of

less than one-third of the citations on their patents. Alcacer and Gittelman (2006) find that

examiners are responsible for 63% of citations on the average patent, and that 40% of patents

have all citations added by the examiners.8 These types of citations might not capture pure8Alcacer et al. (2009) utilise a change in the reporting of US patent data that allows to separate citations

added by the inventor and the examiners to examine the examiners’ behaviour with respect to inventorcitations. In the first case, the patent examiner might add citations that differ in nature from the inven-tor/applicant citations (’gap-filling’). Statistically, the gap-filling scenario would bias estimates of inventorknowledge. In the second case, the examiner might add similar citations (’tracking’). Tracking does not leadto any bias but it may cause standard errors in statistical estimations to be inflated. This raises doubtsabout patent citations as good indicators of knowledge flows. If examiner and inventor citations resembleeach other closely, this suggests that firms and inventors choose their citations with respect to potentialinfrigement and holdup threats and anticipate with some error citations most likely to be added by exam-

13

knowledge spillovers if the inventor was genuinely unaware of that invention.9Fortunately,

our patent data indicate whether the citations was included by the applicant or the patent

examiner. We can thus check the robustness of our results to excluding citations added

by patent examiners.10 Third, inventors and applicants might be strategically referencing

prior art. Citing more prior art will make a patent more valuable in litigation, as it is

much harder to prove a patent is invalid if the patent office has already considered it and

rejected the relevant prior art (Allison et al. (2003)). Most firms employ patent attorneys -

many of whom were formerly patent examiners - to maximise the chances of approval by the

examiner in order to avoid potential infringement and costly holdups. However, inventors

have an incentive not to cite patents unnecessarily as it may reduce their claims to novelty

and therefore affect the scope of the monopoly rights granted by the patent (Hegde and

Sampat (2009); Sampat (2005)). Moreover, not properly referencing priori art can lead to

the invalidation of the patent and is therefore a dangerous strategy.11

2.3 A new measure of spillovers: PatentRank

A potential concern with citation counts is that a citation from an obscure patent is given the

same weight as a citation from a highly-cited work. Hence it is possible that some patents

receive less citations than others but are cited by patents that are themselves more influential

(i.e., more cited themselves). In particular many ground-breaking patents are modestly cited

iners. Moreover, examiners and inventors might exchange information during the application process, andexaminers themselves are prone to biases in favour of citing particular patents. Using the EPO data whichallows to identify the source of the citations since 1979, Criscuolo (2006) attempt to identify the factors thatinfluence whether an observed patent-to-patent citation was added by the applicant/examiner.

9Of course, if the inventor has deliberately omitted to cite a relevant invention, then citations added bypatent examiners actually capture true knowledge spillovers.

10Note that even if the citations was added by the inventor, s/he might have learnt about the citedinvention only after the development of the invention. We have no way to control for this potential isssue.

11“Failure of a person who is involved in the preparation or prosecution of a United States patent applicationto disclose material prior art can result in the patent not issuing, or if issued, being held unenforceableor invalid. As in many instances, the issue of whether prior art is material to patentability can be quitesubjective; it is critical that inventors, assignees, and attorneys be acquainted with the obligations to disclosesuch prior art.” (Silverman (2003a)

14

due to the small size of the scientific community in their area at the time of the publication,

but subsequent patents are themsleves increasingly cited (Maslov and Redner (2009)).

In order to take into account the whole network of patent citations, we apply the random

surfer PageRank algorithm (Page et al. (1999)) to our patent dataset. This algorithm was

originally used by the web search engine Google to help determine the relevance or impor-

tance of a webpage. It does so by analyzing the network of hyperlinks of web pages. The

basic idea is that a webpage is considered important if many other webpages point to it,

or if many webpages point to the webpages that point to it (or both), and so on. To date,

a handful papers have applied this method to rank the importance of patent documents

(Lukach and Lukach (2007); Shaffer (2011)). The resulting PatentRank has the advantage

to readily identify patents that are modestly cited but nevertheless contain ground-breaking

results. It also normalizes the impact of patents from different areas allowing for a more

objective comparison (Maslov and Redner (2009)).

The PatentRank of a patent i is defined as the weighted sum of PatentRanks of all patents

citing i, where the weights depend on the number of citations made by these citing patents.

Therefore, a patent has a high rank if it is cited by many patents with a high rank, and it is

better to be cited by a patent that cites only one patent than by a patent that has a long list

of references. The PatentRank r(i) of patent i is defined according to the following formula

and is computed recursively:12

r(i) =↵

N+ (1� ↵)

X

j✏F (i)

r(j)

B(j)

where N is the total number of patents, F (i) is the set of patents that cite patent i (i.e.

patent i ’s “forward citations”), and B(j) is the number of citations made by patent j (i.e.12The process converges very quickly. In practice we use 50 iterations but the process converges after just

a few iterations.

15

Figure 1: Citation counts and PatentRank

patent j’s number of “backward” citations). The parameter ↵, the damping factor, is used

to avoid sink patents (i.e. patents that are never cited) because sink patents will lead to an

endless loop.13

When constructing the PatentRank, we use the entire population of inventions and their

citations correcting for self-citations by the inventor. We give inventions that are never cited

the smallest PatentRank and rank these PatentRanks to create a PatentRank index. Thus

the higher the PatentRank the greater impact or relevance of the invention. Figure 1shows

that there is a positive correlation overall between the citation count and the PatentRank but

also a vast heterogeneity: many patents have few citations but a high PatentRank and vice

versa. As opposed to citation counts, PatentRank allows us to capture the network centrality

and in particular the influence of a patent. Hence, both indicators are complementary

measures of the intensity of knowledge spillovers.13The mechanism behind the ranking is equivalent to the random-surfer behavior, a person who surfs the

web by randomly clicking links on the visited pages but periodically gets bored and jumps to a randompage altogether. Therefore, when a user is on a web page, she will select one output link randomly withprobability ↵ or will jump to other webpages with probability 1 � ↵. It can be understood as a Markovprocess in which the states are web pages, and the transitions are all equally probable and are the linksbetween webpages.

16

2.4 Exploratory data analysis

The objective of this paper is to compare the extent of knowledge spillovers that arise from

clean and dirty innovations. As shown in table 2, aggregating both sectors together, clean

inventions receive on average 3.40 citations throughout their life time while dirty inventions

receive on average 2.30 citations. This difference is highly statistically significant (see column

3). An obvious problem with this simple comparison is that clean patents are relatively

newer, and hence have had less time to be cited. The average age of clean patents (the

time between the publication year and today) is 22 years as opposed to 27 years for dirty

patents. In order to partly deal with this truncation issue, we look at the number of citations

received within the first five years of the patents’ publication (Hall et al. (2001)). The

difference between the number of citations received by clean and dirty inventions increases:

clean patents receive 74% more citations than dirty patents within their first five years.

Clean inventions also have a significantly higher PatentRank index than dirty inventions.

Looking separately at each technological field, we find that the mean number of citations

and the differences between clean and dirty patents vary across sectors. Inventions in the

transportation sector are more cited overall and have a higher PatentRank. Clean inventions

are more cited and have higher PatentRank than dirty ones in both sectors and this difference

is always significant.

The “innovation flowers” in figure 2 show a network diagram for a random sample of 1000

clean and 1000 dirty innovations where the edges represent citations. This visual represen-

tation of PatentRank highlights the greater PatentRank of clean inventions.

17

Table 2: Mean number of citations and PatentRank

Clean Dirty Diff.

Transport and Electricity

Citations received 3.399 2.295 1.104***(8.256) (5.921) [0.016]

Citations received within 5-years 1.807 1.066 0.741***(4.754) (3.109) [0.009]

PatentRank index 2,335,270 1,920,395 414,874.3***(3,019,924) (2,813,827) [7,354.756]

Transport




Electricity production




Notes: The first two columns report the mean values with standard deviation in parentheses.The last column reports a t-test for the difference in means with standard error in parentheses.*** indicates significance at 0.1% level.

18

Figure 2: Innovation Flowers

(a) Dirty (b) Clean

Notes: The figures visualize innovation spillovoers. We draw a random sample of1000 dirty and 1000 clean innovations corresponding to the nodes in the fig-ures. The edges correspond to backwards citations. An interactive version is underhttp://www.eeclab.org.uk/forcedirect_arx.html?tojson_dirlinks0_1995_15_1000_0.jsonand http://www.eeclab.org.uk/forcedirect_arx.html?tojson_dirlinks0_1995_15_1000_2.json.

19

3 Methods

The simple comparison of reported in the previous section might be confounded by a range

of issues. For example, in recent years there has been a sharp increase in both patenting

and citations. This could reflect a genuine increase in economically relevant spillovers; e.g.

improved IT systems make it easier to learn about previous inventions. However, it could

also reflect legal changes on what consitutes prior art. A similar argument applies when it

comes to innovations from different patent offices. When pooling innovations from different

years there is also a truncation issue: more recent innovations had less time to generate

spillovers than older ones. Most clean innoation activity occured in more recent years.

Clean innovation activities are also more prevalent in some countries than in others. Hence,

such factors might confound any genuine clean effects.

Our strategy to account for this is to estimate a simple count data model of the type

Ci = exp(�Cleani + �Xi + ✏i) (1)

where Ci is the number of citations received by invention i (excluding self-citations) or the

PatentRank index associated to invention i, Cleani is a dummy variable indicating whether

invention i is clean, Xi are controls and ✏i is the error term. Our sample is the population of

clean and dirty patents. Hence, the main coefficient of interest, �, captures the percentage

difference between the number of citations received by clean and dirty patents, all other

things being equal. Given the count data nature of the dependent variable, we estimate

Equation 1 by Poisson pseudo-maximum likelihood. We include a number of control variables

to purge the estimates from as many potential confounding factors as possible which we

discuss in more detail below. However, our basic set of contol variables includes patent

20

office14-by-year-by-sector15 fixed effects. 16

Implicit in the formulation in equaition 1 is the idea that every citation creates economic

value by inspiring a new innovation. It is instructive to express the total spillover value SVi

of an innovation i as follows:

SVi =X

j2J(i).�ji (PVj + SVj) (2)

where J(i) is the set of innovations that could potentially be influenced by innovation i - i.e.

all innovations coming after i - PVj is the private and SVj the spillvoer value of an innovation

j. �ji is the fraction of j’s value that can be attributed to innovation i. By merely looking

at citation counts as in equation 1 we make the assumption that �ij = 1 if j cites i (and zero

otherwise). Moreover we assume that the private values PVj are identical across innovations.

Clearly these are rather strong assumptions although it might be entirely possible that any

error introduced when they are not met is entirely orthorgonal to the question of a clean

spillover gap. To be sure we will relax these assumptions in a number of ways as discussed

in more detail in the following.14For innovations that are patented in multiple countries, we use the patent office where the patent was

patented first.15i.e. energy generation or automotive.16This is implemented following Hausman et al. (1984) by the xtpoisson, fe command in STATA which is

the count data equivalent to the within groups estimator for OLS. Note that Poisson models estimated bypseudo-maximum likelihood can deal with over-dispersion (see Silva and Tenreyro (2006)), so that negativebinomial models offer no particular advantage. In particular, we find the pseudo-fixed effects negativebinomial estimator available in stata (xtnbreg, fe) untrustable, since it does not truly conditions out thefixed effects (only the overdispersion coefficient is assumed to vary across units - see Allison and Waterman(2002); Greene (2007), for more information on this issue). However, as a robustness check we also estimatedEquation (1) using an unconditional negative binomial estimator with patent office, year, month and sectordummies (including a whole range of sector by year by patent office dummies is computationally infeasible)and find very similar results.

21

3.1 Patent Rank

We develop a new method of measuring spillovers in patent data which we call Patent Rank

as it is inspired by the Page-rank algorithm - i.e. the method to rank hyperlinked webpages

that is at the heart of Google’s success as a search engine. This addresses two issues in

particular:

Firstly, it will allways be difficult to attribute and value the specific contribution of prior art.

However, a reasonable alternative to simply assigning �ij = 1 is the inverse of the number

of citations by innovation j:

�ji =1

BCj�i (3)

where BCj are the number of backward citations by innovation j; i.e. if an innovation j cites

a lot of other innovations besides i is likely that the contribution of each invididual i is smaller

than would otherwise be the case. We also include �i < 1 as a scale factor representing the

contribution of spillovers as opposed to other factors such as the firms R&D efforts. Note that

the inverse backward citation weighting can be thought of as observationally equivalent to the

argument made in the introduction: when comapring different technology fields differences

in citations might be driven by a different citation culture where some fields tend to cite

more distant aspects of prior art than others. Hence, if j belongs to citation happy field, i0s

contribution to the value of j is arguably lower. A similar idea is used in the Google pagerank

algorithm. A hyperlink to a web page i from a web page j is considered less valuable if j

hyperlinks a large number of other pages besides i.

Secondly, note that it is unlikely that the spillover values SVj are homogenous across in-

novations. Indeed, our main focus in this study is the heterogeneity of spillovers. By just

counting citations we allow spillover values to be heterogenous on the left hand side of equa-

tion 2 while pretenting that they are uniform on the right hand side of equation 2. However,

22

equation 2 suggests an immediate solution: it implies a system of linear equations (one for

each innovation) which can be solved for the spillover values SVi. Indeed this is what the

page-rank algorithm does as well where citations come in the form of hyperlinks.17

While conceptually, the equation system implied by equation 2 is simple, it raises considerable

computational issues as we are dealing with tens of millions of equations. However, as is

common in Page Rank calculations, the equation system can be solved recursively, rather

than by brute force matrix inversion.18

One further open question concerns the private value of innovations PVj . Clearly, it is very

likely that these vary considerably across innovations. We make a number of assumptions

and examine if this has any impact on our main result of a clean and dirty gap. Firstly, we

assume the same value which we derive from a regression of the impact of a new innovation

on the stock market value of the innovating firm. We discuss these regressions in more detail

below. Secondly, we use common distinctions between high value innovations; e.g. we assign

a non-zero value only to innovations that are “triadic”; i.e. an innovation that is patented in

at least the US, Europe and Japan.

Finally, we need to make assumptions about the values for �i, the contribution of spillovers

as opposed to other factors. In our main results we rely on uniform value of �i = 0.14 which

we derive from an innovation production function type result as reported in Aghion et al.

(2016). The details of this are explained in appendix XXX. However, we also examine the

robustness of our results when using other values.17Although, instead of the private values, the page rank algorithm uses the constant probability of hitting

a web-page randomly - i.e. the inverse of the number of total webpages.18We implemented the procedure in STATA and can make it available as supplementary material to this

paper. We also show in the appendix that none of our modifications to the PageRank algorithm affects itsvalidity.

23

3.2 Evidence from firm values

In addition to using the Patentrank approach we explore the value of different types of

spillover directly by looking at the stock market values of innovating firms. Of course this

restricts our sample to the innovations undertaken by stock listed companies. However, this

will help us detect some potential violations of the assumption we are making above. For

instance, suppose that dirty innovations tend to create knowledge spillovers for innovations

with higher private values on average and the drivers of these higher values are not connected

to easily observed characteristics.19 Or, it could be the case that

To address this we built on the work by Hall et al. (2005), according to which a firm’s knowl-

edge assets are modeled as being accumulated in a continuously ongoing innovative process

where R&D expenditures reflect innovative input, patents record the successful innovations

that can be appropriated by the firm, and citations received by the firm’s patents (forward

citations) measure the relative importance of the patents. We also include citations made

(backward citations) as in Deng (2008) as a proxy of spillovers the firm has received, which

are considered an additional kind of innovative input to direct R&D spendings on the belief

that more knowledge inflows increase the firm’s knowledge stock and may boost the firm’s

R&D productivity. We extend Deng (2008)’s analysis by distinguishing between clean and

dirty backward citations.

Consider Griliches (1981)’s market valuation equation

Vit = qt(Ait + �Kit + �Rit + ⌘Sit) (4)

where Vit denotes firm i’s stock market value in year t, Ait the book value of its physical

assets, and Kit the knowledge assets. qt represents the shadow value of firms’ assets, and

the coefficient b measures the shadow value of knowledge assets relative to physical assets.19Such as the number of patent offices a particular innovation is filed with.

24

� measures the scale effects in the value function and is assumed to be one.

Taking the logarithm, we have the following estimation equation:

logQit = log(

Vit

Ait) = log qt + log(1 + �

Kit

Ait) + "it (5)

where Qit represents Tobin’s q and "it are the prediction errors.

As Deng (2008), we use the following value function to evaluate the firm’s knowledge assets

Kit = f(R&Dit, BCITit,!it) (6)

where R&Ditdenotes the accumulated R&D spendings, BCITit the accumulated backward

citations the firm has made as a proxy of the knowledge inflows received by the firm, and

!it the accumulated idiosyncratic productivity shocks in the firm’s inventive activities. !it

is proxied by the patent / R&D ratio, weighted by the average number of forward citations

the firm’s patents receive over their entire lives (Hall et al. (2005)). This can be viewed as

the knowledge outflow made by the firm.

Taking first-order Taylor expansion of equation 5 yields

Kit = f1 ⇥R&Dit + f2 ⇥ BCITit + f3 ⇥PATit

R&Dit+ f4 ⇥

FCITit

PATit(7)

where PATit and FCITit are firm i’s patent stock and forward citations stock in year t

respectively. Combining equations 5 and 6 leads to

logQit = log qt + log(1 + �1R&Dit

Ait+ �2

BCITit

PATit+ �3

PATit

R&Dit+ �4

FCITit

PATit) + "it (8)

The coefficient �2 represents the value of knowledge flows brought by an additional backward

citation, and �2

�1is a direct measure of the monetary value of knowledge spillovers in terms

25

of R&D equivalent dollar.

We will estimate equation (9) using a non-linear least square including a full set of year and

NACE dummies or firm fixed effects.

26

Table 3: Basic results

(1) (2) (3) (4) (5)

Dep. Variable Citations received Across Technology citations (IPC3) Across Technology (Clean vs dirty) Patent Rank

Clean invention

Patent office-by-year-by-sector no yes yes yes yes

Obs.

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variable is the total number of citations received excluding self-citationsby inventors (columns 1 to 3) and the PatentRank after 20 iterations (columns 4 to 6). All columns are estimated by fixed-effects Poisson pseudo-maximum likelihood.

4 Results

In table 3 we report our basic results for different spillover measures. In columns 1 and 2 we

look at simple citation counts. Columns 1 and 2 look at spillovers across technology classes.

As discussed above, this addresses the concern that the amount of citations that are included

are subject to “cultural” differences in various disciplines. Finally, columns 5 explores our

new patent rank measure. We see that in each case there is an advantage of XX to XX %

for clean innovations.

What could be the drivers and mechanisms behind the subtantial clean spillover gap? In the

following we examine in detail a number of potential reasons.

4.1 Existing and future knowledge stocks

Dirty technologies are still the main driver of most economies. As a consequence knowledge

stock in dirty fields are larger as

27

Table 4: Descriptive statistics on knowledge stocks

(1)

(2)

Cle

anD

irty

Stoc

kbe

fore

wit

hin

IPC

2St

ock

befo

rew

ithi

nIP

C3

Stoc

kbe

fore

with

incl

ean/

dirt

y

Stoc

kaf

ter

wit

hin

IPC

2ye

sye

sSt

ock

afte

rw

ithi

nIP

C3

Stoa

ckaf

ter

with

incl

ean/

dirt

y

Notes:

Table 5: Controlling for future and

(1) (2) (3) (4) (5)

Dep. Variable Citations received Across Technology citations (IPC3) Across Technology (IPC2) Patent Rank

Clean invention

Patent office-by-year-by-sector no yes yes yes yes

Obs.

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variable is the total number of citations received excludingself-citations by inventors (columns 1 to 3) and the PatentRank after 20 iterations (columns 4 to 6). All columns are estimated by fixed-effects Poisson pseudo-maximum likelihood.

28

Figure 3: Clean coefficient between 1950 to 2005 using citations received across IPC3 codes

-1.0

0-0

.50

0.00

0.50

1.00

Cle

an c

oeffi

cien

t

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

4.2 Localized knowledge spillovers

4.3 Firm and inventor characteristics

4.4 Government support

4.5 Quality, Originality, Generality

4.6 The clean spillover gap over time

4.7 Direct measurement of the value of clean vs dirty spillovers

4.8 Clean technology as an emerging field?

29

Table 6: Results by sector

(1) (2) (3) (4)

Sector Transport Electricity Transport Electricity

Dep. var. Citation count PatentRank

Clean invention 0.347*** 0.488*** 0.219*** 0.333***(0.018) (0.023) (0.014) (0.023)

Number of patents -0.068*** -0.047*** -0.048*** -0.019**(0.008) (0.009) (0.006) (0.007)

Family size 0.070*** 0.067*** 0.062*** 0.060***(0.008) (0.004) (0.007) (0.004)

Triadic 0.512*** 0.432*** 0.279*** 0.252***(0.056) (0.050) (0.045) (0.041)

Granted 1.134*** 0.725*** 0.620*** 0.381***(0.034) (0.024) (0.027) (0.017)

Observations 419,959 748,918 419,959 748,918

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variables are the total number of citations received excluding self-citations byinventors in columns 1 and 2 and the PatentRank index in columns 3 and 4. The regres-sions are all estimated by Poisson pseudo-maximum likelihood. The sample includes in-ventions from the transport (columns 1 and 3) and electricity (columns 2 and 4) sectors.All columns include a patent office-by-year and month fixed effects.

30

So far we have focused on the average effect of being a clean invention on the citation

outcome. We now investigate the heterogeneity of the clean premium across the distribution

of citations. Quantile regression techniques are not readily available for count data models,

but we bypass this issue by estimating probit models of the likelihood that a patent falls

within a given percentile of the patent citation distribution (see Chernozhukov et al. (2013)

for a discussion of this issue). We run the following model:

Prob(Citeji = 1) = ↵ + �Cleani + �Xi + ✏i (9)

where Citej i equals one if invention i receives j citations where j varies between 0 (56%

of inventions are never cited) and 479 (the most highly cited invention). Cleani and Xi

are identical to the previous section. Hence the coefficient obtained for Cleani captures

the difference between clean and dirty inventions in the probability of invention i to receive

j citations. Figure 4 shows the coefficient obtained for Cleani and the associated 95%

confidence interval on the number of citations received. We conclude from these results

that (i) clean inventions are always more likely to have a positive citation count than dirty

inventions at all levels of the distribution and (ii) the higher intensity of knowledge spillovers

from clean technologies is even more pronounced for most highly cited patents.

31

Figure 4: Heterogeneity

-1

0

1

2

3

4

5

0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 76 80 84 88 92 96 100

104

109

113

118

122

127

133

144

175

Number of citations received

32

4.9 Localized knowledge spillovers

The existence of localized knowledge spillovers has been widely documented (see Audretsch

and Feldman (2004) for an overview). In one of the earliest papers on this subject, Jaffe

et al. (1993) show that spillovers from research to firms are more intense when the firm is

closer to the institution that generated the research. Jaffe and Trajtenberg (1996, 1999) show

that patent citations tend to occur initially between firms that are close to each other, and

later on spread to a larger geographical area and other countries. Using European patent

data, Maurseth and Verspagen (2002) show that patent citations occur more often between

regions which belong to the same country, same linguistic group and geographical proximity

(see also Peri (2005)). Similar results have been found for energy technologies (see Braun

et al. (2010); Verdolini and Galeotti (2011)).

In our case, clean technologies could generate larger knowledge spillovers than dirty tech-

nologies simply because the clean industry might be more clustered geographically than the

dirty industry. Although we do not have detailed information on the exact localization of

inventors, we do have extensive information on their country of residence. We use this in-

formation to distinguish between national (within-border) and international (cross-border)

citations. We then separately run regressions on these two sets of citation counts.20 For

the PatentRank, we compute a new PatentRank on the pool of national citations and in-

ternational separately. We find that clean inventions exhibit larger national (column 2) and

international (column 3) spillovers. This suggests that clean inventor community transcend

country borders. The clean advantage is larger in terms of domestic spillovers are larger

than international ones.20In the case of collaboration, we weight each citations by the number of inventors from each country

involved in the invention. For example, three inventors working together, one in country A and two incountry B, will count as 1/3 of a citation for country A and 2/3 of a citation for country B.

33

Table 7: Within vs. across-country spillovers

(1) (2) (3)

Dep. var. Citations received Citations received Citations receivedwithin country across country

Clean invention 0.430*** 0.423*** 0.247***(0.014) (0.017) (0.019)

Number of patents -0.057*** -0.057*** -0.081***(0.007) (0.008) (0.006)

Family size 0.073*** 0.062*** 0.066***(0.004) (0.003) (0.004)

Triadic 0.456*** 0.363*** 0.212***(0.036) (0.028) (0.040)

Granted 0.947*** 0.757*** 0.829***(0.031) (0.029) (0.030)

Obs. 1,149,988 1,149,988 1,149,988

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependentvariables are the total number of citations received (column 1), the total number of citations receivedfrom the inventor’s country (column 2), the total number of citations received from all countries ex-cept the invention’s (column 3) corrected for self-citations by inventors. All columns are estimatedby Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

4.10 Public support for R&D

With many clean technologies dependent on policy support of one form or another, the

expansion of clean technologies and its spillovers could be due in part to public investment.

For instance, in 2011 OECD countries spent over 3 billion euros on R&D support to renewable

energy technologies. To control for the government spending level, we include in the first

two columns of table 8 the government spending in clean and dirty technologies within the

transport and electricity sectors. Since we only have information on R&D spending for 28

countries from 1974 onwards, we run the baseline regression for this sample in columns 1, 3

and 5 and the include the government spending in columns 2, 4 and 6. On average, clean

inventions exhibit even larger spillovers than dirty inventions after controlling for government

spending. This effect is driven by the electricity production sector.

Another related concern is that research in clean technologies might come disproportionately

34

from universities rather than private firms. If this is the case, the clean premium might come

from the fact that university patents are more highly cited and more general (Henderson et al.

(1998)). Moreover, the incentive and reward structure within the university system induce

scientists to invest in their reputation by making research publicly available (openness of

the academic community) and make them more willing to recognize the influence of their

predecessors. We control for whether the patent was filed by a university or a firm in the

last two columns of Table 8 with private individuals being the baseline and still find that

clean inventions receive 42% more citations than their dirty counterpart. Finally we run our

baseline regression on the sub-samples of university applicants, firms and individuals. Results

are shown in columns 1, 2, and 3 respectively in table 9. In all three cases, clean inventions

generate more spillovers than their dirty counterparts. Taken together, these results suggest

that public support for R&D is not the driving force behind the clean premium.

35

Table 8: Public spending

(1) (2) (3) (4)

Dep. var. Citations received

Government Spending University

Clean invention 0.493*** 0.507*** 0.421*** 0.423***(0.026) (0.026) (0.014) (0.015)

Number of patents -0.007 -0.006 -0.047*** -0.050***(0.009) (0.009) (0.006) (0.006)

Family size 0.067*** 0.067*** 0.070*** 0.067***(0.004) (0.004) (0.003) (0.003)

Triadic 0.452*** 0.450*** 0.450*** 0.432***(0.046) (0.046) (0.034) (0.034)

Granted 0.689*** 0.688*** 1.005*** 0.992***(0.025) (0.025) (0.031) (0.032)

Government spending 0.034***(0.007)

University 0.429***(0.022)

Firms 0.271***(0.018)

Obs. 496,788 496,788 826,078 826,078

Source: International Energy Agency (2013): Energy Technology Research and Develop-ment Database (Edition: 2013). Mimas, University of ManchesterNotes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citations received excluding self-citations by in-ventors. All columns are estimated by Poisson pseudo-maximum likelihood and includepatent office-by-year and month fixed effects. The samples of columns 1 and 2 includepatent families for which we have government spending, where column 1 is the baselineand column 2 add a control for government spending. The sample of the last two columnsinclude the patent families for which we have university or firm, where column 1 is thebaseline and the column 2 add a control for university and firms.

36

Table 9: University, firms, and private individuals

(1) (2) (3)

Applicant University Firm Individual


Clean invention 0.396*** 0.418*** 0.459***(0.003) (0.016) (0.030)


Family size 0.072*** 0.067*** 0.377***(0.005) (0.003) (0.042)

Triadic 0.152*** 0.454*** -0.870(0.043) (0.035) (0.613)

Granted 0.775*** 1.022*** 0.131***(0.047) (0.032) (0.036)

Obs. 36,186 706,517 75,487

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, ***p<0.001). The dependent variable is the total number of citations re-ceived excluding self-citations by inventors (columns 1 to 3). The sam-ple includes inventions which have universities (column 1 and 4), firms(column 2 and 5), or individuals (column 3 and 6) as applicants. Allcolumns are estimated by Poisson pseudo-maximum likelihood and in-clude patent office-by-year and month fixed effects.

37

4.11 Network effects

Whether guided by “norms of science” (Merton (1957); Small and Griffith (1974)) or self-

interest including personal connections (Leopold (1973); Case and Higgins (2000)), one might

be concerned that inventors working on clean innovation behave systematically differently

from inventors working on dirty innovations. The community of researchers working on clean

technologies could perhaps be smaller and more close-knit. Stuart and Podolny (1996) for

instance argue that there is also a strong social component to a citation. The clean premium

would then represent inventors’ networks rather than true knowledge spillovers. To address

this issue we restrict our sample to inventors who have been working both on clean and dirty

technologies and include inventor fixed effects in our baseline estimations. Our data includes

41,713 such inventors (representing 2.92% of total inventors). Results are presented in table

10. We similarly introduce applicant fixed effects and the results do not change either. The

clean premium remains significant albeit of slightly smaller magnitude. However, this is due

to the different sample as can be seen by comparing columns 1 and 2 and columns 3 and 4

respectively.

38

Table 10: Adding inventor and applicant fixed effect

(1) (2) (3) (4)


Clean invention 0.274*** 0.336*** 0.400*** 0.380***(0.007) (0.011) (0.019) (0.040)

Number of patents -0.096*** -0.081*** -0.038*** -0.067***(0.004) (0.006) (0.008) (0.010)

Family size 0.038*** 0.094*** 0.091*** 0.100***(0.002) (0.006) (0.007) (0.011)

Triadic 0.866*** 0.644*** 0.461*** 0.444***(0.012) (0.026) (0.056) (0.089)

Granted 1.234*** 1.008*** 1.022*** 1.000***(0.007) (0.011) (0.033) (0.046)

Inventor fixed effect no yes no noApplicant fixed effect no no no yes

Obs. 697,192 697,192 435,584 435,584

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citations received excluding self-citations by in-ventors. All columns are estimated by Poisson pseudo-maximum likelihood and includepatent office, sector, year and month fixed effects.

39

4.12 Nature of the citations

There are two important types of citations: references to patent documents that are partic-

ularly close to the new invention, which restrict the claims of the inventor, and references

related to the technological background of the new invention. Therefore citations may reflect

the similarity of inventions rather than the cumulative nature of innovation (Packalen and

Bhattacharya (2012)). To account for the heterogeneous nature of citations, we distinguish

between citations received from inventions in the same technological sector (defined using the

3-digit IPC code as assigned by the patent examiner) and citations received from inventions

in a different technological sector. 21While the former include citations which might merely

reflect similarities between patents, the latter should be closer to true knowledge spillovers.

We then run our baseline regression separately on these two types of citations. Table 11 shows

that clean inventions receive more citations both within and across technological fields, sug-

gesting they do generate larger knowledge spillovers in the economy. The PatentRank index

is computed on the pool of intrasectoral and intersectoral citations separately.

21An important difference between the EPO and the USPTO systems is that in European search reports,cited documents are classified by the patent examiner within a particular citation category according to theirrelevance. When assessing the novelty of patent applications the examiner searches for earlier documentswhich have the same or almost the same features as the patent concerned [Schmoch (1993)].

40

Table 11: Intra vs. inter-sectoral spillovers

(1) (2) (3)

Dep. var. Citations received Intra-sectoral Inter-sectoralcitations citations

Clean invention 0.430*** 0.457*** 0.247***(0.014) (0.015) (0.019)


Family size 0.073*** 0.074*** 0.066***(0.004) (0.004) (0.003)

Triadic 0.456*** 0.487*** 0.212***(0.036) (0.036) (0.040)

Granted 0.947*** 0.963*** 0.829***(0.031) (0.032) (0.030)

Obs. 1,149,988 1,149,988 1,149,988

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, ***p<0.001). s. The dependent variables are the total number of citations (column 1), withina technological field (based on IPC 3 digit code) (column 2), across technological field(column 3) corrected for self-citations by inventors. All columns are estimated by Poissonpseudo-maximum likelihood and include patent office-by-year and month fixed effects.

41

4.13 Generality and Originality

Clean technologies, being relatively newer, might have more opportunities for “fundamental”

research while older dirty technologies might instead be focused on the development of new

applications. If clean technologies have more general applications, this might explain why

they receive more citations and appear to induce larger knowledge spillovers.

In the previous section, clean inventions were found to be more likely to be cited both within

or across their originating technological field. To further investigate the generality of clean

and dirty inventions, we construct a measure of generality based on the Herfindahl index

of concentration introduced by Trajtenberg et al. (1997). It measures the extent to which

the follow-up technical advances (i.e. the citations) are spread across different technological

fields, rather than being concentrated in just a few of them (i.e., they are more likely to have

the characteristics of a General Purpose Technology, see Bresnahan and Trajtenberg (1995);

Popp and Newell (2012)). The generality of a patent is defined in the following way:

Generalityi = 1�niX

j

s2ij (10)

where sij is the percentage of patent citations received by patent family i that belong to

patent class j (definited at 3-digit IPC code), out of ni patent classes.22An originating

patent with generality approaching one receives citations that are very widely dispersed

across patent classes; a generality equal to zero corresponds to the case where all citations

fall into a single class.

Similarly, one might suspect that clean technologies are more original than their dirty coun-

terparts because they are relatively newer. We construct an originality measure using the

same approach as in equation 10 but replacing sij by the percentage of citations made (in-22Specifically, we count the number of citations made by a patent and receiveD by a patent family. This

way we are only capturing citations directly made to an invention as oppose to citations made from onepatent family to another.

42

stead of received) by invention i that belong to patent class j (defined again at 3-digit IPC

code).23Thus, if a patent cites previous patents that belong to a narrow set of technologies

the originality score will be low, whereas citing patents in a wide range of fields would render

a high score.

We carry out regressions using this generality measure as a new outcome variable. Clean

technologies are significantly more general and original in the transport industry while the

opposite in true for the electricity production industry (see Table 12). 24 Adding generality

(column 2), originality (column 3) and finally both measures (column 4) as control in Table 40

confirms the finding of greater knowledge spillovers from clean inventions. Interestingly, the

coefficient is slightly smaller when adding these controls than under the baseline specification

(column 1). This suggests that these measures, particularly the generality measure, explain

(a small) part of the clean premium.

23These measures depend upon the classification system: a finer classification would render higher mea-sures, and conversely for a coarser system. We use 3-digit IPC code as used in Hall et al. (2001)

24Note that there is a potential selection bias here, as patents that have never been cited have no generalitymeasure and are therefore left out of the sample.

43

Table 12: Generality and Originality

(1) (2) (3) (4) (5) (6)

Sector All Transport Electricity All Transport Electricity

Dep. var. Generality measure Originality measure

Clean invention 0.008* 0.047*** -0.034*** -0.003 0.049*** -0.054***(0.003) (0.003) (0.003) (0.004) (0.004) (0.003)

Number of patents -0.047*** -0.081*** -0.024*** -0.050*** -0.086*** -0.027***(0.002) (0.002) (0.001) (0.002) (0.002) (0.001)

Family size 0.012*** 0.011*** 0.012*** 0.008*** 0.007*** 0.007***(0.001) (0.001) (0.001) (0.0004) (0.001) (0.001)

Triadic 0.035*** 0.028*** 0.046*** 0.026*** 0.017*** 0.037***(0.003) (0.004) (0.005) (0.003) (0.003) (0.005)

Granted 0.047*** 0.053*** 0.039*** 0.024*** 0.024*** 0.022***(0.002) (0.002) (0.003) (0.002) (0.003) (0.002)

Observations 515,217 227,678 291,989 382,236 162,919 222,538

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variable is a gener-ality measure (columns 1 to 3) and a originality measure (columns 4 to 6) based on Herfindahl index of concentration.The sample includes patentsin the transport sectors only (colunn 2 and 5), in the electricity sector only (column 3 and6), and in both sectors (columns 1 and 4). All columns are estimated by OLS and include patent office-by-year-by-sectorfixed effects, and month fixed effects.

44

4.14 Clean technologies versus other emerging fields

Technologies that contain a high degree of new knowledge (radical innovations) are likely to

exhibit higher spillover effects than technologies that contain a low degree of new knowledge

(incremental innovations). Clean technologies are new and rather under-developed technolo-

gies. In contrast, the dirty technologies they replace are much more mature and developed.

Therefore research in clean technologies might yield spillovers that are completely different

in scope from research in dirty technologies because they can be considered as radically new

innovations. In order to investigate this assumption, we use several strategies.

4.14.1 Grey innovation

First, we control for the age of the invention’s technological field defined as the time elapsed

since the date of the first appearance of this technological field (defined at the 15-digit IPC

code ) in any patent. Results are reported in column 2 of Table 13. Controlling for the

age of the technology decreases the coefficient obtained for the clean dummy variable. In

order to account for potential non-linearities we further add squared age (column 3) and a

whole range of dummy variables for each percentile of the age distribution (column 4). This

exercise further diminishes the clean coefficient from 0.430 to 0.353, indicating that part of

the clean premium is explained by the relative novelty of the field.

Second, we distinguish between inventions which are radically clean from those which are re-

lated to energy efficiency improvements that make the dirty technology less dirty. So far, our

paper revolves mostly around a distinction between radically clean innovations (e.g. electric

cars, wind turbines) and dirty innovations (e.g. combustion engines, coal power plants). In

the results presented thus far we have included grey innovations in the “dirty” category. We

now identify these inventions and label these “grey” innovations. In tables 14 and 23, we

compare clean inventions with grey inventions (column 2), grey and truly dirty inventions

45

Table 13: Controlling for age of technological field

(1) (2) (3) (4)


Clean invention 0.410*** 0.381*** 0.363*** 0.354***(0.013) (0.013) (0.013) (0.013)


Family size 0.070*** 0.067*** 0.068*** 0.068***(0.004) (0.003) (0.003) (0.003)

Triadic 0.448*** 0.431*** 0.406*** 0.397***(0.035) (0.035) (0.034) (0.034)

Granted 0.939*** 0.929*** 0.917*** 0.912***(0.031) (0.030) (0.030) (0.030)

Age of tech field -0.177*** 0.194***(0.009) (0.034)

Age of tech field^2 -0.023***(0.002)

Age of tech dummies no no no yes

Observations 1,149,237 1,149,237 1,149,237 1,149,237

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citations received, corrected for self-citationsby inventors. All columns are estimated by Poisson pseudo-maximum likelihood and in-clude patent office-by-year and month fixed effects.

46

Table 14: Clean, Grey and True Dirty

(1) (2) (3) (4)

Sample Clean vs. Clean vs. Grey vs. Clean vs.Grey and true Dirty Grey True Dirty True Dirty


Clean/Grey invention 0.430*** 0.191*** 0.307*** 0.502***(0.014) (0.016) (0.016) (0.015)


Family size 0.073*** 0.069*** 0.072*** 0.071***(0.004) (0.007) (0.004) (0.004)

Triadic 0.456*** 0.481*** 0.454*** 0.441***(0.036) (0.055) (0.037) (0.035)

Granted 0.947*** 0.997*** 0.977*** 0.868***(0.031) (0.035) (0.033) (0.027)

Observations 1,149,988 326,942 978,179 1,006,996

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent vari-able is the total number of citations received, corrected for self-citations by inventors. The sample in-cludes clean, grey and truly dirty (column 1), clean and grey (column 2), grey and truly dirty (column 3),and clean and truly dirty (column 4) inventions. All columns are estimated by Poisson pseudo-maximumlikelihood and include patent office-by-year and month fixed effects.

(column 3), and finally clean with truly dirty inventions only (column 4). As a benchmark,

column 1 simply reproduces the results from table6 where grey innovations are included in

the dirty category. This analysis suggests a clear ranking in citations counts: clean technolo-

gies exhibit significantly higher levels of spillovers than grey technologies, which themselves

outperform truly dirty technologies. From a policy perspective, this result implies that radi-

cally clean technologies should receive higher public support than incremental innovation in

dirty technologies.

Third, we compare knowledge spillovers between clean inventions in the transport and elec-

tricity technologies to other radically new technologies, namely IT, biotechnologies, nan-

otechnologies, robots and 3D (see Table 22 for the list of related IPC codes). Results in

table 15 show that clean inventions receive 41% more citations than biotech inventions.

However, clean inventions receive significantly fewer citations than inventions in the IT,

47

Table 15: Spillovers from clean and other new technologies

(1) (2) (3) (4) (5)

Baseline sector IT Biotechs Nano Robot 3D


Clean invention -0.153*** 0.408*** -0.337*** -0.127*** -0.278***(0.029) (0.033) (0.062) (0.042) (0.036)

Number of patents -0.013 -0.160*** -0.031*** -0.039*** -0.037***(0.008) (0.014) (0.008) (0.008) (0.008)

Family size 0.020*** 0.033*** 0.063*** 0.063*** 0.062***(0.003) (0.005) (0.007) (0.007) (0.007)

Triadic 0.574*** 0.663*** 0.525*** 0.550*** 0.528***(0.057) (0.053) (0.070) (0.069) (0.068)

Granted 1.181*** 0.806*** 0.862*** 0.877*** 0.882***(0.065) (0.023) (0.038) (0.036) (0.037)

Observations 1,445,552 403,294 180,441 198,602 185,726

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citations received, corrected for self-citations by inventors.The sample includes all clean patents (transport and electricity) and patents from the following tech-nologies: IT (column 1), bioechs (column 2), nano (column 3), robot (column 4), and 3D (column 5).All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-yearand month fixed effects.

nanotechnology, robot and 3D industries. In tables 40 and 42, we find that clean inventions

are less general and less original than all new technologies apart from nanotechnologies.

Taken together, these results suggest that the relative novelty of clean technologies might

explain why they exhibit larger spillovers. Looking at the coefficients obtained for the clean

invention variable, it is interesting to note that knowledge spillovers from clean technologies

appear comparable to those in the IT sector, which has been behind the third industrial

revolution.

Fourth, we compare the previous sample (clean transport, clean electricity, IT, biotech,

nano, robots and 3D) to all other inventions. Figure 5 plots the coefficient of the dirty (in

black), grey (in grey), clean (in green) and radically new technologies (in orange). Clean

transport and electricity exhibit larger spillovers than the average invention. In terms of

relative ranking, the clean transport and clean electricity are positioned between their dirty

48

Figure 5: Clean, grey, dirty, and radically new technologies vs. all other technologies-Citations count

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

1.2

Electricitydirty

Electricitygrey

Car dirty

Biotechs

Electricity clean

Car grey

Car clean

Robot IT

Nano

3D

counterparts and radically new technologies.

Fifth, we restrict the sample of radically new technologies (IT, biotechs, nano, and robots)

and compare clean and dirty inventions within these technologies. While clean inventions

within the IT and the biotechs technologies still exhibit larger knowledge spillovers, there is

no clean advantage within the nano and robot sectors.

Finally, in an attempt to find a dirty yet radically new technology, we compare knowledge

spillovers between clean electricity production technologies and carbon capture and storage

technologies (CCS) in table 17. The clean advantage disappears when considering simple

patent counts and PatentRank, suggesting it is not because they are clean that clean tech-

nologies generate larger knowledge spillovers.

49

Table 16: Comparing spillovers from clean and dirty within new technologies

(1) (2) (3) (4)

Sector IT Biotechs Nano Robot


Clean invention 0.222* 0.609** 0.313 0.677(0.091) (0.053) (0.211) (0.525)

Number of patents -0.012 -0.257*** -0.169*** -0.051(0.008) (0.016) (0.044) (0.047)

Family size 0.020*** 0.033*** 0.109*** 0.104***(0.003) (0.005) (0.018) (0.014)

Triadic 0.547*** 0.583*** 0.268* 0.387***(0.055) (0.056) (0.136) (0.113)

Granted 1.220*** 0.699*** 0.961*** 1.005***(0.072) (0.031) (0.145) (0.053)

Observations 1,270,842 227,100 1,481 22,266

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, ***p<0.001). The dependent variable is the total number of citations received, correctedfor self-citations by inventors. The sample includes patents from the following tech-nologies: IT (column 1), bioechs (column 2), nano (column 3), robot (column 4),and 3D (column 5). All columns are estimated by Poisson pseudo-maximum likeli-hood and include patent office-by-year and month fixed effects.

50

Table 17: Spillovers from clean and CCS technologies


Clean invention -0.083*(0.034)

Number of patents 0.037***(0.010)

Family size 0.065***(0.006)

Triadic 0.477***(0.062)

Granted 0.681***(0.030)

Observations 106,700

Notes: Robust standard errors, p-values in paren-theses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citationsreceived, corrected for self-citations by inventors.The sample includes clean electricity productioninventions and CO2 Capture and Storage technol-ogy. All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

51

5 Monetary value of knowledge spillovers

5.1 Market value equation

In order to quantify the economic value of knowledge spillovers, in particular clean knowledge

spillovers, we estimate a market valuation equation using firm-level data. Following Hall et al.

(2005), a firm’s knowledge assets are modeled as being accumulated in a continuously ongoing

innovative process in which R&D expenditures reflect innovative input, patents record the

successful innovations that can be appropriated by the firm, and citations received by the

firm’s patents (forward citations) measure the relative importance of the patents. We also

include citations made (backward citations) as in Deng (2008) as a proxy of the knowledge

flows the firm has received, which are considered an additional kind of innovative input to

direct R&D spendings on the belief that more knowledge inflows increase the firm’s knowledge

stock and may boost the firm’s R&D productivity. We extend Deng (2008)’s analysis by

further distinguishing between clean and dirty backward citations to capture knowledge

spillovers from clean and dirty technologies.

Consider Griliches (1981)’s market valuation equation

Vit = qt(Ait + �Kit + �Rit + ⌘Sit) (11)

where Vit denotes firm i’s stock market value in year t, Ait the book value of its physical

assets, and Kit the knowledge assets. qt represents the shadow value of firms’ assets, and

the coefficient b measures the shadow value of knowledge assets relative to physical assets.

� measures the scale effects in the value function and is assumed to be one.

Taking the logarithm, we have the following estimation equation:

logQit = log(

Vit

Ait) = log qt + log(1 + �

Kit

Ait) + "it (12)

52

where Qit represents Tobin’s q and "it are the prediction errors.

As Deng (2008), we use the following value function to evaluate the firm’s knowledge assets

Kit = f(R&Dit, BCITit,!it) (13)

where R&Ditdenotes the accumulated R&D spendings, BCITit the accumulated backward

citations the firm has made as a proxy of the knowledge inflows received by the firm, and

!it the accumulated idiosyncratic productivity shocks in the firm’s inventive activities. !it

is proxied by the patent / R&D ratio, weighted by the average number of forward citations

the firm’s patents receive over their entire lives (Hall et al. (2005)). This can be viewed as

the knowledge outflow made by the firm.

Taking first-order Taylor expansion of equation 5 yields

Kit = f1 ⇥R&Dit + f2 ⇥ BCITit + f3 ⇥PATit

R&Dit+ f4 ⇥

FCITit

PATit(14)

where PATit and FCITit are firm i’s patent stock and forward citations stock in year t

respectively. Combining equations 5 and 6 leads to

logQit = log qt + log(1 + �1R&Dit

Ait+ �2

BCITit

PATit+ �3

PATit

R&Dit+ �4

FCITit

PATit) + "it (15)

The coefficient �2 represents the value of knowledge flows brought by an additional backward

citation, and �2

�1is a direct measure of the monetary value of knowledge spillovers in terms

of R&D equivalent dollar.

We will estimate equation (9) using a non-linear least square including a full set of year and

NACE dummies or firm fixed effects.

53

5.2 Firm level data

For this purpose, we combine the PATSTAT database with the ORBIS database which

contains firm-level information such as R&D expenses, market value, and total number of

assets. The analysis focuses on 10,299 firms from 2001 to 2011 for which we can match

both datasets and identify each firm’s patents along with the citations (both backward and

forward) associated to them.

We calculate the stock of R&D (patents) as the accumulated past R&D expenditures (the

number of patents) subject to an annual deprecitation rate assumed to be a constant 10%.

As previously we measure patents and their citations at the patent family level. The stock of

backward citations is measured taking into account the age of the patent and then aggregate

them over the firm’s patent portfolio each year subject to the annual depreciation of 15%.

We remove self-citations among these backward citations. The stock of forward citations

measures the relative importance of a firm’s portfolio. Given the truncation issues associated

to the tine lag in observing forward citations, we limit our sample up to 2011 and scale

citations taking into account the average citations across publication years, patent offices,

sector and citation year. We finally aggregate these scaled forward citations subject to annual

depreciation. A table of descriptive statistics can be found in table 18.

5.3 Estimating the market value of clean knowledge spillover

Table 19 shows a significant positive monetary value for knowledge spillovers. The first

three columns include a full set of year and NACE dummies, while the last three columns

incorporate firm fixed effects. The coefficient estimates of R&D/assets and forward cita-

tions/patents are positive and significant. A one-percentage point increase in R&D/assets

ratio leads to a 0.4% appreciation in the firm value. A rise in the average quality of the

firm’s patent portfolio also raises the firm’s market value – if every patent receives one more

54

Table 18: Descriptive Statistics, 2001-2008, N = 29,154

Mean Median Min Max Std. Dev.Market Value ($M) 2.860 0.208 0.001 512 13.9Total assets ($M) 3.526 0.319 0.00007 798 17.5Tobin’s Q 1.148 0.695 0.045 15.517 1.476Profits/Losses ($M) 0.358 0.027 0 83.4 1.852Operating revenue ($M) 2.862 0.285 0 460 12.3Number of employees 8424.873 1266 1 550000 28321.2R&D stock ($M) 0.191 0.005 0 37.5 1.225R&D stock ($M) for 20,745 obs. with R&D>0 0.266 0.017 1 37.5 1.439D(R&D) = 0 0.282 0 0 1 0.450Patent stock 254.033 4.892 0 88013.93 2252.876R&D stock / Total assets 0.1585 0.021 0 209.654 1.459Patent stock / R&D stock 0.014 0.001 0 34.163 0.355Fwd citation stock / Patent stock 4.428 2.149 0 222 7.668Bwd citation stock / Patent stock 2.928 1.635 0 77.004 4.133Clean bwd citation stock / Patent stock 0.038 0 0 11.746 0.316Dirty bwd citation stock / Patent stock 0.059 0 0 8.299 0.316Other bwd citation stock / Patent stock 2.831 1.538 0 76.482 4.099

forward citation over their entire lives, the firm’s value will rise between 0.07% and 0.1%.

Adding knowledge spillovers means that one extra backward citation per patent makes the

firm about 0.07% more valuable. The amount of appreciation is no longer significant once

we control for firm fixed effects. Columns 3 and 6 distinguishes among the type of spillovers.

We see that knowledge spillovers from clean and other technologies are positive although less

significant when including firm fixed effects. Of particular interest is the monetary value of

clean spillovers. One extra clean backward citation per patent makes the firm about 0.13%

more valuable, and the amount of appreciation is even larger to 0.15% when we include firm

fixed effects. Given the fact that patents recive on average of 2.899 citations in our sample,

our results translate to an average

We conduct a number of robustness checks. First,we investigate our results by sector (NACE

codes) and firm country. We find that our results are mainly driven by firms in the man-

ufacturing industry and in the United States (see table 50). Second, we add a number of

55

Table 19: Estimation of Tobin’s Q equation

(1) (2) (3) (4) (5) (6)

Dep. var. ln Tobin’s Q

ln (R&D/assets) 0.441*** 0.431*** 0.432*** 0.144*** 0.145*** 0.145***(0.029) (0.029) (0.029) (0.037) (0.037) (0.037)

ln (Patent/R&D) 0.020 0.040 0.040 -0.030 -0.027 -0.027(0.043) (0.045) (0.045) (0.023) (0.024) (0.024)

ln (Fwd citations/patent) 0.101*** 0.048*** 0.047*** 0.076*** 0.064*** 0.064***(0.006) (0.009) (0.009) (0.013) (0.014) (0.014)

ln (Bwd cites/patent) 0.078*** 0.030**(0.010) (0.014)

ln (Bwd clean cites/ patent) 0.118*** 0.149**(0.039) (0.066)

ln (Bwd dirty cites/ patent) 0.050 0.038(0.034) (0.042)

ln (Bwd other cites/ patent) 0.074*** 0.009(0.010) (0.014)

D(R&D=0) 0.080*** 0.080*** 0.080*** 0.007 0.007 0.013(0.011) (0.011) (0.011) (0.015) (0.015) (0.015)

Year + NACE Yes Yes Yes No No NoFirm FE No No No Yes Yes Yes

Observations 23,654 23,654 23,654 23,659 23,659 23,659

Notes: Clustered standard errors in parentheses (* p<0.1, ** p<0.05, *** p<0.01). The dependent variable is the lnTobin’s Q defined as the stock market value over the book value of physical assets. We restrict the sample to patentsapplied for between 2000 and 2008. All columns are estimated by OLS. Columns 1, 2, and 3 include a complete set ofyear and NACE dummies while columns 4, 5, and 6 include firm fixed effects.

56

variables to control for the quality of the patents. We use the average characteristics such as

the family size, generality and originality of cited patents, and the PatentRank of the patents

held by the firm each year (see table 51). Third, we base our patent and citation counts

on granted patents only. Patent counts may suffer from biases due to truncation because

only a fraction of patents are eventually granted and, if they are granted, there is a lag

between application and granted dates (see table 52). All these robustness checks confirm

the conclusion found in the main table of results.

57

6 Discussion and conclusion

In this paper we compare the relative intensity of knowledge spillovers from clean and dirty

technologies. To measure knowledge spillovers, we use a rich dataset of 3 million citations

received by over a million inventions patented in the automobile and electricity production

sectors. This analysis is crucial to answer the question of whether clean technologies warrant

higher subsidies than dirty ones. Our results unambiguously show that clean technologies

induce larger knowledge spillovers than their dirty counterparts. Moreover, we provide a

measure of the economic values of knowledge externalities of relevant clean technologies. We

find that evidence of larger monetary value associated to knowledge spillovers from clean

technology. We conduct a large number of sensitivity tests and the findings are remark-

ably robust. In particular, as depicted by the innovation flowers, this result is confirmed

when using a completely novel methodology to measure knowledge spillovers that does not

only count immediate forward citations but takes into account the whole network of patent

citations.

We explore five potential explanations for our findings. First, we find no evidence that

the clean industry is more geographically clustered. Second, differential citations behaviors

among scientists involved in clean technologies cannot fully explain the clean advantage.

Third, we find no evidence that government spending cannot account for clean premium.

Fourth, we examine the generality and originality features of clean inventions. We find that

clean inventions in the automobile industry are more general (i.e. they are cited by a wider

range of technological fields) and more original. However, clean inventions in the electricity

production industry are less general and less original. Finally, we compare clean inventions

to other radically new inventions such as IT, biotechnologies and nanotechnologies. We

conclude that clean inventions seem to benefit from early returns to scale and steep learning

curves. Interestingly we observe that knowledge spillovers from clean technologies appear

58

comparable in scope to those in the IT sector.

Our results have two important policy implications. Firstly, the larger knowledge spillovers

from clean technologies uncovered in this study justify higher subsidies for clean R&D or

specific R&D programs for clean technologies, in addition to implicit support for clean R&D

trough climate policies such as carbon taxation. Radically new clean technologies should

receive higher public support than research activities targeted at improving on the existing

dirty technologies.25 However, such specific support could equally be justified for a range of

other emerging areas, such as nanotechnologies or IT. This recommendation has been made

in the past, for instance by Hart (2008) or Acemoglu et al. (2012) but it is the first time to our

knowledge that it is substantiated by robust empirical evidence.26While a first best policy

scenario would suggest a combination of emissions pricing and R&D subsidies specifically

targeted at clean technologies, in times of tight government budgets it might be difficult to

achieve the necessary subsidy levels. There might also be concerns over governments’ ability

to channel funds to R&D projects with the highest potential either because of information

asymmetry or because of political interference. In this case our results would support a

second best policy with more stringent emission pricing and regulation that would otherwise

be the case (see for example Gerlagh et al. (2009); Hart (2008); Kverndokk et al. (2004);

Kverndokk and Rosendahl (2007)).

Secondly, our results lend support to the idea that a redirection of innovation from dirty to

clean technologies induced by environmental or climate policies can lead to higher growth25Importantly, our results suggest that the relative support to clean R&D should grow over time. In-

cidentally, in a recent working paper Daubanes et al. (2013) show that gradual rise in subsidies to cleanR&D activities causes a less rapid extraction of fossil resources, because it enhances the long-run resourceproductivity.

26Interestingly, statistics in OECD countries show that there is higher public R&D spending in cleantechnologies than in dirty ones. A look at the International Energy Agency’s R&D expenditures data revealsthat between 2000 and 2012, OECD countries have spent 198 million euros on dirty cars and 18 billion euroson dirty energy but 327 million euros on clean cars (65% more than dirty cars) and 25 billion euros on cleanenergy (35% more than dirty energy). However, these numbers do not include subsidies to private cleanR&D, which is also warranted in a first best policy setting.

59

in the short and medium run. This can happen if the larger spillover effects from clean

technologies exceed any negative growth effects from more stringent regulation. Our results

however suggest that the potential growth effects of environmental policies very much depend

on the type of displacement being induced by increasing support for clean technologies. If

clean innovation crowds out dirty innovation, as shown by Aghion et al. (2012) for the

transport industry, there is scope for medium run growth effects. If innovation in other

emerging areas is crowded out, such effects are less likely. At any rate, one should keep in

mind that higher spillovers are only a necessary but not a sufficient condition for growth

effects from green policies.

Our work can be extended in several directions. First, it would be interesting to investigate

how knowledge spillovers affect firms’ decisions to invest in radical innovation (clean technolo-

gies) or in incremental innovation (less dirty technologies), and how they respond to R&D

subsidies targeted at clean technologies. Second, an interesting direction is to understand

the spatial pattern of knowledge diffusion for clean technologies, including the transfer of

knowledge across borders, in particular between developed and developing countries. Third,

we could use micro data to estimate the impact of knowledge spillovers from clean and dirty

technologies on firms’ productivity. These parameters are crucial to empirically validate the

potential impact of green policies on economic growth.

Radically new clean innovations that require consumers to substitute to a different prod-

uct (e.g. electric vehicles replacing internal combustion engine propelled vehicles) or more

incremental innovations that improve the energy efficiency of current (dirty) products (for

example fuel efficiency technologies for combustion engines). We label this latter category

as “grey” inventions. From a policy point of view an important question is whether to give

priority to clean or grey innovation in order to mitigate climate change (see Aghion et al.

(2012), for a further discussion on this issue).

60

References

Acemoglu, D., Aghion, P., Bursztyn, L., and Hemous, D. (2012). The environment and

directed technical change. American Economic Review, 102(1):131–66.

Aghion, P., Dechezleprêtre, A., Hemous, D., Martin, R., and Van Reenen, J. (2012). Carbon

taxes, path dependency and directed technical change: Evidence from the auto industry.

NBER Working Papers 18596, National Bureau of Economic Research, Inc.

Aghion, P. and Howitt, P. (1992). A model of growth through creative destruction. Econo-

metrica, 60 (2):323–351.

Aghion, P. and Howitt, P. (1996). Research and development in the growth process. Journal

of Economic Growth, 1(1):49–73.

Aghion, P. and Howitt, P. (1998). Endogenous growth theory. MIT press.

Alcacer, J. and Gittelman, M. (2006). Patent citations as a measure of knowledge flows: The

influence of examiner citations. The Review of Economics and Statistics, 88(4):774–779.

Alcacer, J., Gittelman, M., and Sampat, B. (2009). Applicant and examiner citations in us

patents: An overview and analysis. Research Policy, 38(2):415–427.

Allison, J. R., Lemley, M. A., Moore, K. A., and Trunkey, R. D. (2003). Valuable patents.

Geo. Lj, 92:435.

Allison, P. D. and Waterman, R. P. (2002). Fixed–effects negative binomial regression mod-

els. Sociological methodology, 32(1):247–265.

Ambec, S. and Barla, P. (2006). Can environmental regulations be good for business? an

assessment of the porter hypothesis. Energy studies review, 14(2):1.

61

Ambec, S., Cohen, M. A., Elgie, S., and Lanoie, P. (2013). The porter hypothesis at 20: can

environmental regulation enhance innovation and competitiveness? Review of Environ-

mental Economics and Policy, 7(1):2–22.

Audretsch, D. and Feldman, M. (2004). Knowledge spillovers and the geography of innova-

tion. Handbook of regional and urban economics, 4:2713–2739.

Bjørner, T. B. and Mackenhauer, J. (2013). Spillover from private energy research. Resource

and Energy Economics.

Braun, F., Schmidt-Ehmcke, J., and Zloczysti, P. (2010). Innovative activity in wind and

solar technology: Empirical evidence on knowledge spillovers using patent data.

Breschi, S. and Lissoni, F. (2005). Knowledge networks from patent data. In Handbook of

quantitative science and technology research, pages 613–643. Springer.

Bresnahan, T. F. and Trajtenberg, M. (1995). General purpose technologies ’engines of

growth’? Journal of econometrics, 65(1):83–108.

Caballero, R. and Jaffe, A. (1993). How high are the giants’ shoulders: An empirical assess-

ment of knowledge spillovers and creative destruction in a model of economic growth. In

NBER Macroeconomics Annual 1993, Volume 8, pages 15–86. MIT press.

Case, D. O. and Higgins, G. M. (2000). How can we investigate citation behavior? a study

of reasons for citing literature in communication. Journal of the American Society for

Information Science, 51(7):635–645.

Chernozhukov, V., Fernández-Val, I., and Melly, B. (2013). Inference on counterfactual

distributions. Econometrica, 81(6):2205–2268.

62

Cockburn, I., Kortum, S., and Stern, S. (2003). Are all patent examiners equal? examiners,

patent characteristics, and litigation outcomes. Patents in the knowledge-based economy,

35.

Collins, P. and Wyatt, S. (1988). Citations in patents to the basic research literature.

Research Policy, 17(2):65–74.

Criscuolo, P. (2006). The’home advantage’effect and patent families. a comparison of oecd

triadic patents, the uspto and the epo. Scientometrics, 66(1):23–41.

Daubanes, J., Grimaud, A., and Rougé, L. (2013). Green paradox and directed technical

change: The effects of subsidies to clean R&D. Technical report, Institut d’Économie

Industrielle (IDEI), Toulouse.

Deng, Y. (2008). The value of knowledge spillovers in the us semiconductor industry. Inter-

national Journal of Industrial Organization, 26(4):1044–1058.

Dernis, H. and Khan, M. (2004). Triadic patent families methodology. Technical report,

OECD Publishing.

Fischer, C. and Newell, R. (2008). Environmental and technology policies for climate miti-

gation. Journal of environmental economics and management, 55(2):142–162.

Fischer, C., Newell, R., and Preonas, L. (2013). Environmental and technology policy options

in the electricity sector: Interactions and outcomes. Working paper.

Gerlagh, R., Kverndokk, S., and Rosendahl, K. E. (2009). Optimal timing of climate change

policy: Interaction between carbon taxes and innovation externalities. Environmental and

resource Economics, 43(3):369–390.

Goulder, L. H. and Schneider, S. H. (1999). Induced technological change and the attrac-

tiveness of co2 abatement policies. Resource and Energy Economics, 21(3):211–253.

63

Greaker, M. and Heggedal, T.-R. (2012). A comment on the environment and directed

technical change. Technical report.

Greene, W. H. (2007). Fixed and random effects models for count data. Working paper,

Department of Economics, Stern School of Business, New York University, New York.

Griliches, Z. (1981). Market value,< i> r&d</i>, and patents. Economics letters, 7(2):183–

187.

Griliches, Z. (1992). The search for R&D spillovers. Technical report, National Bureau of

Economic Research.

Griliches, Z., Hall, B. H., and Pakes, A. (1991). R&D, patents, and market value revisited:

is there a second (technological opportunity) factor? Economics of Innovation and new

technology, 1(3):183–201.

Grossman, G. and Helpman, E. (1991). Innovation and growth in the world economy.

Grupp, H. (1998). Foundations of the economics of innovation: theory, measurement, and

practice. E. Elgar.

Grupp, H., Münt, G., and Schmoch, U. (1996). Assessing different types of patent data

for describing high-technology export performance. Innovation, patents and technological

strategies, pages 271–287.

Guellec, D. and Van Pottelsberghe de la Potterie, B. (2004). From R&D to productivity

growth: Do the institutional settings and the source of funds of R&D matter? Oxford

Bulletin of Economics and Statistics, 66(3):353–378.

Hall, B. H., Jaffe, A. B., and Trajtenberg, M. (2001). The nber patent citation data file:

Lessons, insights and methodological tools. Technical report, National Bureau of Economic

Research.

64

Hall, B. H., Jaffe, A. B., and Trajtenberg, M. (2005). Market value and patent citations.

RAND Journal of economics, pages 16–38.

Harhoff, D., Scherer, F. M., and Vopel, K. (2003). Citations, family size, opposition and the

value of patent rights. Research Policy, 32(8):1343–1363.

Hart, R. (2004). Growth, environment and innovation - a model with production vintages and

environmentally oriented research. Journal of Environmental Economics and Management,

48(3):1078–1098.

Hart, R. (2007). Can environmental regulations boost growth? In Sustainable Resource Use

and Economic Dynamics, pages 53–70. Springer.

Hart, R. (2008). The timing and balance of policies for CO2 abatement when technological

change is endogenous. Journal of Environmental Economics and Management, 55:194–212.

Hausman, J., Hall, B. H., and Griliches, Z. (1984). Econometric models for count data with

an application to the patents- R&D relationship. Econometrica, 52(4):909–38.

Hegde, D. and Sampat, B. (2009). Examiner citations, applicant citations, and the private

value of patents. Economics Letters, 105(3):287–289.

Henderson, R., Jaffe, A. B., and Trajtenberg, M. (1998). Universities as a source of com-

mercial technology: a detailed analysis of university patenting, 1965–1988. Review of

Economics and Statistics, 80(1):119–127.

Jaffe, A. B., Newell, R. G., and Stavins, R. (2005). A tale of two market failures: Technology

and environmental policy. Ecological Economics, 54(2):164–174.

Jaffe, A. B. and Trajtenberg, M. (1996). Flows of knowledge from universities and federal

labs: Modeling the flowof patent citations over time and across institutional and geographic

boundari. NBER Working Papers 5712, National Bureau of Economic Research, Inc.

65

Jaffe, A. B. and Trajtenberg, M. (1999). International knowledge flows: Evidence from

patent citations. Economics of Innovation and New Technology, 8(1-2):105–136.

Jaffe, A. B., Trajtenberg, M., and Fogarty, M. S. (2000). Knowledge spillovers and patent

citations: Evidence from a survey of inventors. NBER/Sloan˜, page 21.

Jaffe, A. B., Trajtenberg, M., and Henderson, R. (1993). Geographic localization of

knowledge spillovers as evidenced by patent citations. Quarterly Journal of Economics,

108(3):577–598.

Jones, C. I. (2005). Growth and ideas. Handbook of Economic Growth, 1:1063–1111.

Keller, W. (2004). International technology diffusion. Journal of Economic Literature,

42(3):752–782.

Kverndokk, S. and Rosendahl, K. E. (2007). Climate policies and learning by doing: Impacts

and timing of technology subsidies. Resource and Energy Economics, 29(1):58–82.

Kverndokk, S., Rosendahl, K. E., and Rutherford, T. F. (2004). Climate policies and induced

technological change: Which to choose, the carrot or the stick? Environmental and

Resource Economics, 27(1):21–41.

Lanjouw, J. and Mody, A. (1996). Innovation and the international diffusion of environmen-

tally responsive technology. Research Policy, 25(4):549–571.

Lanjouw, J. O. and Schankerman, M. (1999). The quality of ideas: measuring innovation

with multiple indicators. Technical report, National Bureau of Economic Research.

Lanjouw, J. O. and Schankerman, M. (2004). Patent quality and research productivity:

Measuring innovation with multiple indicators. The Economic Journal, 114(495):441–465.

Leopold, A. (1973). Games scientists play. BioScience, 23(10):590–594.

66

Lukach, R. and Lukach, M. (2007). Ranking USPTO patent documents by importance using

random surfer method (pagerank). Available at SSRN 996595.

Maslov, S. and Redner, S. (2009). Promise and pitfalls of extending google’s pagerank

algorithm to citation networks. Mimeo, Division of Material Science, U.S. Department of

Energy.

Maurseth, P. and Verspagen, B. (2002). Knowledge spillovers in europe: a patent citations

analysis. The Scandinavian journal of economics, 104(4):531–545.

Merton, R. K. (1957). Priorities in scientific discovery: a chapter in the sociology of science.

American sociological review, pages 635–659.

Michel, J. and Bettels, B. (2001). Patent citation analysis. a closer look at the basic input

data from patent search reports. Scientometrics, 51(1):185–201.

Mohr, R. D. (2002). Technical change, external economies, and the porter hypothesis. Jour-

nal of Environmental economics and management, 43(1):158–168.

Neuhoff, K. (2005). Large-scale deployment of renewables for electricity generation. Oxford

Review of Economic Policy, 21(1):88–110.

Noailly, J., Shestalova, V., et al. (2013). Knowledge spillovers from renewable energy tech-

nologies lessons from patent citations. Technical report, CPB Netherlands Bureau for

Economic Policy Analysis.

Otto, V. M., Löschel, A., and Dellink, R. (2007). Energy biased technical change: a cge

analysis. Resource and Energy Economics, 29(2):137–158.

Otto, V. M., Löschel, A., and Reilly, J. (2008). Directed technical change and differentiation

of climate policy. Energy Economics, 30(6):2855–2878.

67

Packalen, M. and Bhattacharya, J. (2012). Words in patents: Research inputs and the value

of innovativeness in invention. Technical report, National Bureau of Economic Research.

Page, L., Brin, S., Motwani, R., and Winograd, T. (1999). The pagerank citation ranking:

bringing order to the web.

Peri, G. (2005). Determinants of knowledge flows and their effect on innovation. Review of

Economics and Statistics, 87(2):308–322.

Popp, D. and Newell, R. (2012). Where does energy R&D come from? examining crowding

out from energy R&D. Energy Economics, 34(4):980–991.

Popp, D., Newell, R., and Jaffe, A. (2009). Energy, the environment, and technological

change. NBER Working Papers 14832, National Bureau of Economic Research, Inc.

Porter, M. E. and Van der Linde, C. (1995). Toward a new conception of the environment-

competitiveness relationship. The journal of economic perspectives, 9(4):97–118.

Ricci, F. (2007a). Channels of transmission of environmental policy to economic growth: A

survey of the theory. Ecological Economics, 60(4):688–699.

Ricci, F. (2007b). Environmental policy and growth when inputs are differentiated in pollu-

tion intensity. Environmental and Resource Economics, 38(3):285–310.

Romer, P. M. (1990). Endogenous technological change. Journal of political Economy, pages

S71–S102.

Sampat, B. (2005). Determinants of patent quality: an empirical analysis. Columbia Univ.,

New York.

Scherer, F. M. (1965). Firm size, market structure, opportunity, and the output of patented

inventions. The American Economic Review, 55(5):1097–1125.

68

Schmoch, U. (1993). Tracing the knowledge transfer from science to technology as reflected

in patent indicators. Scientometrics, 26(1):193–211.

Schmookler, J. (1966). Invention and economic growth, volume 26. Harvard University Press

Cambridge, MA.

Schneider, S. H. and Goulder, L. H. (1997). Achieving low-cost emissions targets. Nature,

389(6646):13–14.

Shaffer, M. J. (2011). Entrepreneurial Innovation: Patent Rank and Marketing Science. PhD

thesis, Washington State University.

Silva, J. S. and Tenreyro, S. (2006). The log of gravity. The Review of Economics and

Statistics, 88(4):641–658.

Silverman, A. (2003a). Duty to disclose prior art to the united states patent and trademark

office. JOM, 55:64–64.

Silverman, A. B. (2003b). Duty to disclose prior art to the united states patent and trademark

office. JOM Journal of the Minerals, Metals and Materials Society, 55(10):64–64.

Small, H. and Griffith, B. C. (1974). The structure of scientific literatures i: Identifying and

graphing specialties. Science studies, pages 17–40.

Smulders, S. and De Nooij, M. (2003). The impact of energy conservation on technology and

economic growth. Resource and Energy Economics, 25(1):59–79.

Smulders, S. and Withagen, C. (2012). Green growth–lessons from growth theory. World

Bank Policy Research Working Paper, (6230).

Stuart, T. E. and Podolny, J. M. (1996). Local search and the evolution of technological

capabilities. Strategic Management Journal, 17(S1):21–38.

69

Thompson, P. and Fox-Kean, M. (2005). Patent citations and the geography of knowledge

spillovers: A reassessment. American Economic Review, pages 450–460.

Trajtenberg, M. (1990). A penny for your quotes: patent citations and the value of innova-

tions. The Rand Journal of Economics, pages 172–187.

Trajtenberg, M., Henderson, R., and Jaffe, A. (1997). University versus corporate patents:

A window on the basicness of invention. Economics of Innovation and New Technology,

5(1):19–50.

Van Pottelsberghe, B., Denis, H., and Guellec, D. (2001). Using patent counts for cross-

country comparisons of technology output. Technical report, ULB–Universite Libre de

Bruxelles.

Veefkind, V., Hurtado-Albir, J., Angelucci, S., Karachalios, K., and Thumm, N. (2012). A

new EPO classification scheme for climate change mitigation technologies. World Patent

Information, 34(2):106–111.

Verdolini, E. and Galeotti, M. (2011). At home and abroad: An empirical analysis of

innovation and diffusion in energy technologies. Journal of Environmental Economics and

Management, 61(2):119–134.

Appendix

70

Table 20: Patent classification codes - Transport

CLEAN

B60K 1 Arrangement or mounting of electrical propulsion unitsB60K 6 Arrangement or mounting of hybrid propulsion systems comprising

electric motors and internal combustionB60L 3 Electric devices on electrically-propelled vehicles for safety purposes:

Monitoring operating variables e.g. speed, deceleration, power consumptionB60L 7 Dynamic electric regenerative brakingB60L 11 Electric propulsion with power supplied within the vehicleB60L 15 Methods, circuits, or devices for controlling the traction-motor speed

of electrically-propelled vehiclesB60R 16 Electric or fluid circuits specially adapted for vehicles and not otherwise provided forB60S 5 Supplying batteries to, or removing batteries formB60W 10 Conjoint control of vehicles sub-units of different type or different functionB60W 20 Control systems specially adapted for hybrid vehiclesH01M Fuel cells

GREY

F02M 39/71 Fuel injection apparatusF02M 3/02-05 Idling devices for carburettors preventing flow of idling fueF02M 23 Apparatus for adding secondary air to fuel-air mixtureF02M 25 Engine-pertinent apparatus for adding non-fuel substances or small quantities

of secondary fuel to combustion-air, main fuel, or fuel-air mixtureF02D 41 Electric control of supply of combustion mixture or its constituentsF02B 47/06 Methods of operating engines involving adding non-fuel substances or

anti-knock agents to combustion air, fule, or fuel-air mixtures of engines,the substances including non-airborne oxygen

DIRTY

F02B Internal-combustion piston engines; combustion engines in generaF02D Controlling combustion enginesF02F Cylinders, pistons, or casings for combustion engines;

arragement of sealings in combustion enginesF02M Supplying combustion engines with combustiles mixtures or constituents thereofF02N Starting of combustion enginesF02P Ignition (other than compression ignition) for internal-combustion engines

71

Table 21: Patent classification codes - Electricity Production

CLEAN

Y02E10 Energy generation through renewable energy sourcesY02E30 Energy generation of nuclear originE02B9/08 Tide or wave power plantsF03B13/10-26 Submerged units incorporating electric generators or motors characterized

by using wave or tide energyF03D Wind motorsF03G4 Devices for producing mechanical power from geothermal energyF03G6 Devices for producing mechanical power from solar energyF03G7/05 Ocean thermal energy conversionF24J2 Use of solar heat, e.g. solar heat collectorsF24J3/08 Production or use of heat, not derived from combustion using geothermal heatF26B3/28 Drying solid materials or objects by processes involving the application of

heat by radiation, e.g. from the sun

GREY

Y02E50 Technologies for the production of fuel of non-fossil originY02E20/10 Combined combustionY02E20/12 Heat utilisation in combustion or incineration of wasteY02E20/14 Combined heat and power generationY02E20/16 Combined cycle power plant, or combined cycle gas turbineY02E20/18 Integrated gasification combined cycleY02E20/30 Technologies for a more efficient combustion or heat usageY02E20/32 Direct C02 mitigationY02E20/34 Indirect C02 mitigation, by acting on non C02 directly related matters of

the process, more efficient use of fuelsY02E20/36 Heat recovery other than air pre-heating

DIRTY

C10G1 Production of liquid hydrocarbon mixtures from oil-shale, oil-sand, or non-melting solidcarbonaceous or similar materials, e.g. wood, coal, oil-sand, or the like B03B

C10L1 FuelC10J Production of fuel gases by carburetting air or other gasesE02B Hydraulic engineeringF01K Steam engine plans; steam accumulators; engine plants not otherwise provided for;

engines using special working fluids or cyclesF02C Gas-turbine plants; air intakes for jet-propulsion plants; controlling fuel supply

in air-breathing jet-propulsion plantsF22 Steam generationF23 Combustion apparatus; combustion processesF24J Production or use of heat not otherwise provided forF27 Furnaces; kilns; ovens; retortsF28 Heat exchange in general 72

Table 22: Patent classification codes - Radically New Technologies

3D

H04N 13 Stereoscopic television systems

IT

G06 Computing; Calculating; CountingG10L Speech Analysis or Synthesis; Speech Recognition; Speech or Voice Processing; Speech or Audio Coding or DecodingG11C Static Stores

(not G06Q)Data Processing Systems or Methods; Specially Adapted for Administrative, Commercial, Financial,Managerial, Supervisory or Forecasting purposes; Systems or Methods Specially Adapted for Administrative,Commercial, Financial, Managerial, Supervisory or Forecasting purposes, not otherwise provided for

Biotechs

C07G Compounds of unknown constitutionC07K PeptidesC12M Apparatus for Enzymology or MicrobiologyC12N Micro-organisms or enzymes; compositions thereofC12P Fermentation or Enzyme-using Processes to Synthesise a desired chemical

compound or composition or to separate optical isomers from a racemic mixtureC12Q Measuring or Testing Processes Involving Enzymes or Micro-Organisms;

Compositions or test papers therefor; Processes of preparing such compositions;Condition responsive control in microbiological or enzymological processes

C12R Processes using micro-organisms(not A61K) Preparations for Medical, Dental, or Toilet Purposes

Nano

B82 Nano-technology

Robot

B25J 9 Programme-controlled manipulators

73

Patent examples

Figure 6: Patent example

74


75


76

Figure 9: The gap in knowledge spillovers between 1950 and 2005 using PatentRank

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005

PatentRank Results

77

Figure 10: The gap in knowledge spillovers between 1950 and 2005 using PatentRank

(a) No future patents

-0.2

00.

000.

200.

400.

60C

lean

coe

ffici

ent

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

(b) Future patents

-0.2

00.

000.

200.

400.

60C

lean

coe

ffici

ent

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

Figure 11: The gap in knowledge spillovers between 1950 and 2005 using PatentRank within5 years window


-0.2

00.

000.

200.

400.

60C

lean

coe

ffici

ent

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

(b) Future patents

-0.2

00.

000.

200.

400.

60C

lean

coe

ffici

ent

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

78

Figure 12: The gap in knowledge spillovers between 1950 and 2005 using PatentRank acrossIPC3 codes


-0.2

00.

000.

200.

400.

60C

lean

coe

ffici

ent

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

(b) Future patents-0

.20

0.00

0.20

0.40

0.60

Cle

an c

oeffi

cien

t

1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005Year

79

Table 23: Clean, Grey and True Dirty

(1) (2) (3) (4)


Dep. var. PatentRank index


Number of patents -0.031*** -0.006 -0.084*** -0.029***(0.005) (0.008) (0.004) (0.005)

Family size 0.067*** 0.059*** 0.065*** 0.065***(0.003) (0.006) (0.004) (0.003)

Triadic 0.241*** 0.278*** 0.238*** 0.240***(0.025) (0.045) (0.028) (0.026)

Granted 0.491*** 0.520*** 0.508*** 0.456***(0.021) (0.022) (0.022) (0.019)

Observations 1,149,988 326,942 978,179 1,006,996

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependentvariable is the PatentRank index. The sample includes clean, grey and truly dirty (column 1), cleanand grey (column 2), grey and truly dirty (column 3), and clean and truly dirty (column 4) inventions.All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year andmonth fixed effects.

Figure 13: Clean, grey, dirty, and radically new technologies vs. all other technologies -PageRank index

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Electricitydirty

ElectricitygreyCar

dirtyBiotechs

Electricity clean

Car grey

Car clean

RobotIT

Nano3D

80

Table 24: Within vs. across-country spillovers

(1) (2) (3)

Dep. var. PatentRank PatentRank PatentRankfor “national” citations for “international” citations

Clean invention 0.292*** 0.285*** 0.361***(0.014) (0.017) (0.013)


Family size 0.067*** 0.062*** 0.073***(0.003) (0.003) (0.003)

Triadic 0.241*** 0.240** 0.331***(0.026) (0.020) (0.033)

Granted 0.491*** 0.435*** 0.731***(0.021) (0.016) (0.028)

Obs. 1,149,988 1,149,988 1,149,988

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variablesare the PatentRank index (column 1), the PatentRank index on the pool of national citations (column 2),and the PatentRank index on the pool of international citations (column 3). All columns are estimated byPoisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

81

Table 25: Government spending

(1) (2) (3) (4) (5) (6)

Sample All Transport Electricity


Clean invention 0.345*** 0.353*** 0.153** 0.149** 0.339** 0.347**(0.028) (0.026) (0.052) (0.052) (0.028) (0.026)

Government spending 0.022*** -0.020 0.021**(0.007) (0.155) (0.006)

Number of patents 0.012 0.013 -0.040** -0.040** 0.013 0.014(0.008) (0.008) (0.015) (0.015) (0.008) (0.008)

Family size 0.060*** 0.059*** 0.057*** 0.057*** 0.059*** 0.059***(0.004) (0.004) (0.015) (0.015) (0.004) (0.004)

Triadic 0.285*** 0.284*** 0.391*** 0.394*** 0.274*** 0.273***(0.037) (0.037) (0.076) (0.076) (0.037) (0.037)

Granted 0.360*** 0.360*** 0.534*** 0.535*** 0.359*** 0.358***(0.017) (0.017) (0.032) (0.032) (0.017) (0.017)

Obs. 497,439 497,439 16,719 16,719 489,531 489,531

Source: International Energy Agency (2013): Energy Technology Research and Development Database (Edition:2013). Mimas, University of ManchesterNotes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variable is thePatentRank index. The sample includes clean and dirty inventions from the transport sector (columns 3 and 4),electricity sector (columns 5 and 6) and both transport and electricity sectors (columns 1 and 2) . All columns areestimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

82

Table 26: Government spending

(1) (2) (3) (4)

Sample Transport Electricity


Clean invention 0.253** 0.253*** 0.483*** 0.497***(0.077) (0.079) (0.026) (0.026)

Government spending -0.001 0.032***(0.033) (0.007)

Number of patents -0.070*** -0.070*** -0.006 -0.005(0.020) (0.020) (0.009) (0.009)

Family size 0.054*** 0.054*** 0.066*** 0.066***(0.012) (0.012) (0.004) (0.004)

Triadic 0.474*** 0.474*** 0.447*** 0.445***(0.093) (0.094) (0.046) (0.047)

Granted 0.776*** 0.776*** 0.696*** 0.695***(0.055) (0.055) (0.026) (0.026)

Obs. 16,703 16,703 488,896 488,896

Source: International Energy Agency (2013): Energy Technology Research and Devel-opment Database (Edition: 2013). Mimas, University of ManchesterNotes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citations received excluding self-citations byinventors. The sample includes clean and dirty inventions from the transport sector(columns 3 and 4), electricity sector (columns 5 and 6) and both transport and electric-ity sectors (columns 1 and 2) . All columns are estimated by Poisson pseudo-maximumlikelihood and include patent office-by-year and month fixed effects.

83

Table 27: University and Firms

(1) (2)


Clean invention 0.293*** 0.298***(0.013) (0.013)

Number of patents -0.019*** -0.022***(0.004) (0.004)

Family size 0.063*** 0.060***(0.003) (0.003)

Triadic 0.237*** 0.229***(0.024) (0.024)

Granted 0.561*** 0.552***(0.021) (0.021)

University 0.276***(0.014)

Firms 0.206***(0.011)

Obs. 826,078 826,078

Notes: Robust standard errors in parentheses (*p<0.05, ** p<0.01, *** p<0.001). The dependent vari-able is the total number of citations received excludingself-citations by inventors (columns 1 and 2) and thePatentRank index (columns 3 and 4). All columns areestimated by Poisson pseudo-maximum likelihood andinclude patent office-by-year and month fixed effects.

84

Table 28: University, firms, and private individuals

(1) (2) (3)

Applicant University Firm Individual


Clean invention 0.311*** 0.290*** 0.331***(0.016) (0.014) (0.019)


Family size 0.056*** 0.060*** 0.289***(0.005) (0.003) (0.036)

Triadic 0.096** 0.239*** -0.614(0.030) (0.025) (0.340)

Granted 0.411*** 0.571*** 0.088***(0.030) (0.022) (0.025)

Obs. 36,186 706,517 75,487

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01,*** p<0.001). The dependent variable is the total number of citationsreceived excluding self-citations by inventors (columns 1 to 3) and thePatentRank index (columns 4 to 6). The sample includes inventionswhich have universities (column 1 and 4), firms (column 2 and 5), or in-dividuals (column 3 and 6) as applicants. All columns are estimated byPoisson pseudo-maximum likelihood and include patent office-by-yearand month fixed effects.

85

Table 29: Adding inventor and inventor fixed effect

(1) (2) (3) (4)


Clean invention 0.216*** 0.259*** 0.274*** 0.272***(0.004) (0.006) (0.017) (0.027)

Number of patents -0.028*** -0.023*** -0.002 -0.024***(0.002) (0.003) (0.007) (0.008)

Family size 0.027*** 0.077*** 0.082*** 0.085***(0.002) (0.004) (0.006) (0.009)

Triadic 0.598*** 0.405*** 0.250*** 0.254***(0.009) (0.015) (0.042) (0.056)

Granted 0.721*** 0.572*** 0.562*** 0.574***(0.005) (0.007) (0.024) (0.025)

fixed effect no inventor no applicant

Obs. 697,192 697,192 435,584 435,584

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001).The dependent variable is the total number of citations received excluding self-citations by inventors. All columns are estimated by Poisson pseudo-maximum like-lihood and include patent office, sector, year and month fixed effects.

Table 30: Intra vs. inter-sectoral spillovers

(1) (2) (3)

Dep. var. PatentRank PatentRank PatentRankintra-sectoral inter-sectoral

Clean invention 0.292*** 0.336*** 0.248***(0.014) (0.016) (0.016)


Family size 0.067*** 0.067*** 0.068***(0.003) (0.003) (0.003)

Triadic 0.241*** 0.246*** 0.259***(0.026) (0.025) (0.025)

Granted 0.491*** 0.456*** 0.521***(0.021) (0.021) (0.017)

Obs. 1,149,988 1,149,988 1,149,988

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01,*** p<0.001). The dependent variables are PatentRank index (column 1), Paten-tRank index on citations within their own technological field (based on IPC 3digit code) (column 2), and the PatentRank index on citations across acrosstechnological field (column 3). All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

86

Table 31: Generality and originality as controls

(1) (2) (3) (4)


Clean invention 0.193*** 0.179*** 0.193*** 0.178***(0.007) (0.007) (0.007) (0.007)

Number of patents -0.010*** 0.019*** -0.003 0.016***(0.002) (0.002) (0.002) (0.002)

Family size 0.026*** 0.022*** 0.025*** 0.023***(0.001) (0.001) (0.001) (0.001)

Triadic 0.130*** 0.110*** 0.127*** 0.111***(0.007) (0.006) (0.007) (0.006)

Granted 0.245*** 0.203*** 0.240*** 0.204***(0.010) (0.010) (0.010) (0.010)

Generality 0.628*** 0.663***(0.010) (0.010)

Originality 0.127*** -0.097***(0.006) (0.008)

Obs. 281,978 281,978 281,978 281,978

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001).The dependent variable is the total number of citations received, corrected for self-citations by inventors. All columns are estimated by Poisson pseudo-maximumlikelihood and include patent office-by-year and month fixed effects.

87

Table 32: Controlling for age of technological field

(1) (2) (3) (4)


Clean invention 0.283*** 0.267*** 0.257*** 0.247***(0.013) (0.013) (0.013) (0.012)


Family size 0.065*** 0.063*** 0.063*** 0.063***(0.003) (0.003) (0.003) (0.003)

Triadic 0.236*** 0.227*** 0.210*** 0.202***(0.025) (0.025) (0.025) (0.025)

Granted 0.487*** 0.480*** 0.474*** 0.470***(0.021) (0.021) (0.020) (0.020)

Age of tech field -0.117*** 0.233***(0.006) (0.014)

Age of tech field^2 -0.023***(0.001)

Age of tech dummies no no no yes

Observations 1,149,237 1,149,237 1,149,237 1,149,237

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the PatentRank index. All columns are estimated by Poissonpseudo-maximum likelihood and include patent office-by-year and month fixed effects.

88

Table 33: Spillovers from clean and other new technologies

(1) (2) (3) (4) (5)

Baseline sector IT Biotechs Nano Robot 3D


Clean invention -0.039 0.131*** -0.249*** -0.096* -0.120***(0.028) (0.023) (0.040) (0.043) (0.018)

Number of patents -0.031*** -0.029*** 0.023*** 0.014 0.018*(0.005) (0.006) (0.008) (0.078) (0.008)

Family size 0.017*** 0.029*** 0.052*** 0.053*** 0.052***(0.003) (0.004) (0.006) (0.006) (0.006)

Triadic 0.421*** 0.435*** 0.329*** 0.337*** 0.333***(0.050) (0.042) (0.055) (0.054) (0.055)

Granted 0.604*** 0.413*** 0.441*** 0.443*** 0.448***(0.040) (0.017) (0.025) (0.025) (0.025)

Observations 1,445,552 403,294 180,441 198,602 185,726

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the PatentRank index. The sample includes all clean patents (transport andelectricity) and patents from the following technologies: IT (column 1), biotechs (column 2), nano(column 3), robot (column 4), and 3D (column 5). All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

89

Table 34: Comparing spillovers from clean and dirty within new technologies

(1) (2) (3) (4) (5)

Sector IT Biotechs Nano Robot 3D


Clean invention 0.129* 0.422*** 0.189 0.349 0.290(0.053) (0.067) (0.100) (0.325) (0.461)

Number of patents -0.037*** -0.074*** 0.033 -0.062** -0.080***(0.006) (0.005) (0.023) (0.023) (0.014)

Family size 0.017*** 0.028*** 0.070*** 0.088*** 0.056***(0.003) (0.004) (0.013) (0.009) (0.011)

Triadic 0.401*** 0.406*** 0.341*** 0.261*** 0.305**(0.049) (0.041) (0.070) (0.057) (0.060)

Granted 0.624*** 0.342*** 0.424*** 0.443*** 0.571***(0.044) (0.019) (0.075) (0.042) (0.044)

Observations 1,270,842 227,100 1,481 22,266 9,359

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the PatentRank. The sample includes patents from the following technolo-gies: IT (column 1), bioechs (column 2), nano (column 3), robot (column 4), and 3D (column 5).All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

90

Table 35: Spillovers from clean and CCS technologies


Clean invention 0.045(0.023)

Number of patents 0.057***(0.010)

Family size 0.055***(0.005)

Triadic 0.271***(0.047)

Granted 0.338***(0.019)

Observations 106,700

Notes: Robust standard errors, p-values in paren-theses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the total number of citationsreceived, corrected for self-citations by inventors.The sample includes clean electricity productioninventions and CO2 Capture and Storage technol-ogy. All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixed effects.

91

Table 36: Clean, Grey and true Dirty - Transport

(1) (2) (3) (4)





Family size 0.070*** 0.070*** 0.081*** 0.065***(0.008) (0.011) (0.010) (0.007)

Triadic 0.521*** 0.483*** 0.474*** 0.488***(0.056) (0.071) (0.059) (0.055)

Granted 1.134*** 1.122*** 1.173*** 1.046***(0.034) (0.041) (0.036) (0.032)

Observations 419,959 207,524 345,313 287,469

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The de-pendent variable is the total number of citations received, corrected for self-citations by inventors. Thesample includes clean, grey and truly dirty (column 1), clean and grey (column 2), grey and truly dirty(column 3), and clean and truly dirty (column 4) inventions all in the transport sector. All columnsare estimated by Poisson pseudo-maximum likelihood and include patent office-by-year and month fixedeffects.

Clean, Grey and True Dirty

92

Table 37: Clean, Grey and true Dirty - Transport

(1) (2) (3) (4)





Family size 0.062*** 0.059*** 0.074*** 0.057***(0.007) (0.010) (0.007) (0.006)

Triadic 0.279*** 0.281*** 0.219*** 0.284***(0.046) (0.057) (0.040) (0.046)

Granted 0.620*** 0.599*** 0.637*** 0.588***(0.027) (0.027) (0.029) (0.024)

Observations 419,959 207,524 345,313 287,469

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The de-pendent variable is PatentRank index. The sample includes clean, grey and truly dirty (column 1), cleanand grey (column 2), grey and truly dirty (column 3), and clean and truly dirty (column 4) inventions allin the transport sector. All columns are estimated by Poisson pseudo-maximum likelihood and includepatent office-by-year and month fixed effects.

93

Table 38: Clean, Grey and true Dirty - Electricity

(1) (2) (3) (4)




Number of patents -0.047*** 0.042*** -0.114*** -0.044***(0.009) (0.011) (0.007) (0.009)

Family size 0.067*** 0.070*** 0.066*** 0.067***(0.004) (0.004) (0.004) (0.004)

Triadic 0.432*** 0.416*** 0.396*** 0.438***(0.050) (0.051) (0.046) (0.050)

Granted 0.725*** 0.660*** 0.738*** 0.727***(0.024) (0.029) (0.026) (0.025)

Observations 748,918 120,752 647,541 733,859

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The de-pendent variable is the total number of citations received, corrected for self-citations by inventors. Thesample includes clean, grey and truly dirty (column 1), clean and grey (column 2), grey and truly dirty(column 3), and clean and truly dirty (column 4) inventions all in the electricity production sector.All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year andmonth fixed effects.

94

Table 39: Clean, Grey and true Dirty - Electricity

(1) (2) (3) (4)



Clean/Grey invention 0.333*** 0.046 0.287*** 0.342***(0.023) (0.028) (0.013) (0.023)

Number of patents -0.019*** 0.062*** -0.073*** -0.015*(0.007) (0.010) (0.004) (0.007)

Family size 0.060*** 0.058*** 0.061*** 0.061***(0.004) (0.003) (0.004) (0.005)

Triadic 0.252*** 0.228*** 0.226*** 0.256***(0.000) (0.038) (0.037) (0.042)

Granted 0.381*** 0.331*** 0.393*** 0.382***(0.017) (0.018) (0.018) (0.017)

Observations 748,918 120,752 647,541 733,859

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The de-pendent variable is the PatentRank index. The sample includes clean, grey and truly dirty (column 1),clean and grey (column 2), grey and truly dirty (column 3), and clean and truly dirty (column 4) in-ventions all in the electricity production sector. All columns are estimated by Poisson pseudo-maximumlikelihood and include patent office-by-year and month fixed effects.

95

Table 40: Generality and originality as controls

(1) (2) (3) (4)


Clean invention 0.365*** 0.332*** 0.363*** 0.332***(0.012) (0.012) (0.012) (0.012)

Number of patents -0.044*** 0.007 -0.025*** 0.006(0.005) (0.006) (0.005) (0.005)

Family size 0.043*** 0.039*** 0.041*** 0.039***(0.002) (0.002) (0.002) (0.002)

Triadic 0.296*** 0.264*** 0.287*** 0.264***(0.014) (0.013) (0.014) (0.013)

Granted 0.673*** 0.591*** 0.659*** 0.592***(0.023) (0.021) (0.022) (0.021)

Generality 1.149*** 1.164***(0.019) (0.019)

Originality 0.371*** -0.036*(0.015) (0.015)

Obs. 281,978 281,978 281,978 281,978

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001).The dependent variable is the total number of citations received, corrected for self-citations by inventors. All columns are estimated by Poisson pseudo-maximumlikelihood and include patent office-by-year and month fixed effects.

96

Table 41: Comparing the generality of clean and other new technologiess

(1) (2) (3) (4) (5)


Dep. var. Originality measure

Clean invention -0.050*** -0.059*** 0.009 -0.130*** -0.184***(0.004) (0.004) (0.018) (0.004) (0.006)

Number of patents -0.070*** -0.033*** -0.049*** -0.051*** -0.051***(0.002) (0.002) (0.002) (0.002) (0.002)

Family size 0.003*** 0.002*** 0.005*** 0.005*** 0.005***(0.0004) (0.0003) (0.001) (0.001) (0.001)

Triadic 0.010*** 0.005 0.014*** 0.015*** 0.014**(0.002) (0.002) (0.005) (0.004) (0.004)

Granted 0.020*** -0.003 0.029*** 0.027*** 0.027***(0.002) (0.003) (0.003) (0.003) (0.003)

Observations 520,978 155,701 59,651 67,115 62,559

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependentvariable is the originality measure. The sample includes all clean inventions (automobile and elec-tricity production sectors) and inventions from the following technologies: IT (column 1), biotechs(column 2), nano (column 3), robot (column 4), and 3D (column 5). All columns are estimated byOLS and include patent office-by-year and month fixed effects.

97

Table 42: Comparing the generality of clean and other new technologiess

(1) (2) (3) (4) (5)


Dep. var. Generality measure

Clean invention -0.047*** -0.052*** 0.009 -0.126*** -0.204***(0.004) (0.004) (0.022) (0.004) (0.006)

Number of patents -0.063*** -0.034*** -0.048*** -0.050*** -0.050***(0.002) (0.002) (0.003) (0.003) (0.003)

Family size 0.004*** 0.003*** 0.008*** 0.008*** 0.008***(0.0005) (0.0003) (0.001) (0.001) (0.001)

Triadic 0.013*** 0.016*** 0.022*** 0.023*** 0.020***(0.002) (0.003) (0.005) (0.005) (0.004)

Granted 0.022*** 0.022*** 0.041*** 0.038*** 0.039***(0.002) (0.002) (0.003) (0.003) (0.003)

Observations 723,257 207,073 94,437 103,972 98,461

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependentvariable is the generality measure. The sample includes all clean patents (automobile and electric-ity production sectors) and patents from the following technologies: IT (column 1), biotechs (column2), nano (column 3), robot (column 4), and 3D (column 5). All columns are estimated by OLS andinclude patent office-by-year and month fixed effects.

98

Table 43: Five-year window

(1) (2) (3) (4) (5) (6)

Sector All Transport Electricity All Transport Electricity

Dep. var. Citations received within 5-year window PatentRank within 5-year window

Clean invention 0.382*** 0.284*** 0.474*** 0.210*** 0.140*** 0.248***(0.021) (0.025) (0.034) (0.015) (0.014) (0.026)

Number of patents -0.038*** -0.055*** -0.023* -0.038*** -0.059*** -0.022**(0.008) (0.001) (0.010) (0.005) (0.006) (0.007)

Family size 0.075*** 0.070*** 0.063*** 0.069*** 0.062*** 0.059***(0.003) (0.001) (0.007) (0.003) (0.008) (0.006)

Triadic 0.508*** 0.557*** 0.515*** 0.306*** 0.354*** 0.346***(0.043) (0.070) (0.068) (0.003) (0.053) (0.053)

Granted 1.005*** 1.181*** 0.756*** 0.581*** 0.693*** 0.473***(0.040) (0.054) (0.035) (0.024) (0.036) (0.022)

Observations 1,162,220 419,959 748,918 1,162,220 419,959 748,918

Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variable is the totalnumber of citations received within a five-year period after the publication year, corrected for self-citations by inventors(columns 1 to 3) and the PatentRank index on the sample of citations within five years (columns 4 to 6). The sample in-cludes patents which have cited clean or dirty technologies in the automobile sector (columns 2 and 4), electricity sector(columns 3 and 6), and both (colunns 1 and 4). All columns are estimated by Poisson pseudo-maximum likelihood andinclude patent office-by-year-by-sector fixed effects, and month fixed effects.

Robustness Checks

Five years window

As in section 2.4 we look at the number of citations received within a five-year window to at

least partially overcome the truncation bias that is due to the fact that we observe citations

for only a portion of the life of an invention, with the duration of that portion varying across

patent cohorts (see Table 43). The coefficients obtained for the clean dummy barely change.

99

Discarding citations

We discard citations added by patent examiners in Table 44.27By restricting the citation

counts to the ones made by the applicant only, we address the concern that patent citations

added by examiners might not capture actual knowledge spillovers. The results obtained

when all sectors are pooled together barely change but the only noticeable difference is

that the clean dummy is no longer significant on the fuel sector when citations added by

examiners are excluded. Jaffe and Trajtenberg (1999)) find that patent assigned to the same

firm are more likely to cite each other. We therefore correct for self-citations at the level of

the applicant (the firm or the individual who filed the patent) rather than at the level of

individual inventors in Table 45. The results don’t change qualitatively.

27Note that we restrict the sample to patent offices for which distinction between citation added by patentexaminer or applicant is made.

100

Tabl

e44

:C

itatio

nsm

ade

byap

plic

ants

only

(1)

(2)

(3)

(4)

(5)

(6)

Sect

orA

llTr

ansp

ort

Ele

ctri

city

All

Tran

spor

tE

lect

rici

ty

Dep

.va

r.C

itat

ions

rece

ived

excl

.ci

tati

ons

adde

dby

pate

ntex

amin

erPa

tent

Ran

kin

dex

excl

.ci

tati

ons

adde

dby

pate

ntex

amin

er

Cle

anin

vent

ion

0.62

4***

0.58

2***

0.65

9***

0.04

1***

0.01

30.

085*

**(0

.018

)(0

.025

)(0

.026

)(0

.009

)(0

.011

)(0

.009

)N

umbe

rof

pate

nts

-0.0

10-0

.036

***

0.00

7-0

.016

***

-0.0

16**

*-0

.018

***

(0.0

08)

(0.0

11)

(0.0

11)

(0.0

02)

(0.0

03)

(0.0

03)

Fam

ilysi

ze0.

070*

**0.

070*

**0.

060*

**0.

012*

**0.

017*

**0.

008*

**(0

.003

)(0

.007

)(0

.004

)(0

.001

)(0

.002

)(0

.001

)Tr

iadi

c0.

516*

**0.

537*

**0.

564*

**0.

072*

**0.

049*

**0.

098*

**(0

.042

)(0

.064

)(0

.066

)(0

.008

)(0

.009

)(0

.011

)G

rant

ed1.

144*

**1.

164*

**1.

101*

**0.

102*

**0.

097*

**0.

105*

**(0

.025

)(0

.035

)(0

.031

)(0

.007

)(0

.009

)(0

.012

)

Obs

erva

tion

s1,

162,

220

419,

950

748,

918

1,16

2,22

041

9,95

074

8,91

8

Notes:

Rob

ust

stan

dard

erro

rsin

pare

nthe

ses

(*p<

0.05

,**

p<0.

01,*

**p<

0.00

1).

The

depe

nden

tva

riab

leis

the

tota

lnum

ber

ofci

tati

ons

rece

ived

,exc

ludi

ngci

tati

ons

mad

eby

the

pate

ntex

amin

ers

and

corr

ecte

dfo

rse

lf-ci

tati

ons

byin

vent

ors

(col

umns

1to

3)an

dth

ePat

entR

ank

inde

xon

the

sam

ple

ofci

tati

ons

excl

udin

gth

eon

esm

ade

byth

epa

tent

exam

iner

s(c

olum

s4

to6)

.T

hesa

mpl

eon

lyin

clud

espa

tent

office

sfo

rw

hich

the

info

rmat

ion

onw

heth

erci

tati

ons

are

adde

dby

the

appl

ican

tor

the

pate

ntex

amin

erar

eav

aila

ble.

The

sam

ple

incl

udes

pate

nts

whi

chha

veci

ted

clea

nor

dirt

yte

chno

logi

esin

the

auto

mob

ilese

ctor

(col

umns

2an

d4)

,el

ectr

icity

prod

ucti

onse

ctor

(col

umns

3an

d6)

,in

both

(col

unns

1an

d4)

.A

llco

lum

nsar

ees

tim

ated

byPoi

sson

pseu

do-m

axim

umlik

elih

ood

and

incl

ude

pate

ntoffi

ce-b

y-ye

ar-b

y-se

ctor

fixed

effec

ts,an

dm

onth

fixed

effec

ts.

101

Tabl

e45

:E

xclu

ding

self-

cita

tions

atap

plic

ant

leve

l

(1)

(2)

(3)

(4)

(5)

(6)

Sect

orA

llTr

ansp

ort

Ele

ctri

city

All

Tran

spor

tE

lect

rici

ty

Dep

.va

r.C

itat

ions

rece

ived

corr

ecte

dfo

rse

lf-ci

tati

ons

atap

plic

ant

leve

lPa

tent

Ran

kin

dex

corr

ecte

dfo

rse

lf-ci

tati

ons

atap

plic

ant

leve

l

Cle

anin

vent

ion

0.36

3***

0.35

1***

0.35

0***

0.16

4***

0.16

1***

0.15

5***

(0.0

13)

(0.0

22)

(0.0

16)

(0.0

07)

(0.0

11)

(0.0

10)

Num

ber

ofpa

tent

s-0

.040

***

-0.0

30**

*-0

.047

***

0.00

040.

002

0.00

1(0

.005

)(0

.007

)(0

.007

)(0

.003

)(0

.004

)(0

.004

)Fa

mily

size

0.04

4***

0.04

6***

0.03

8***

0.03

3***

0.03

3***

0.03

0***

(0.0

03)

(0.0

05)

(0.0

03)

(0.0

02)

(0.0

03)

(0.0

02)

Tria

dic

0.21

8***

0.25

0***

0.21

4**

0.11

4***

0.10

3***

0.13

5**

(0.0

24)

(0.0

39)

(0.0

30)

(0.0

13)

(0.0

19)

(0.0

15)

Gra

nted

0.45

6***

0.54

9***

0.38

8***

0.20

1***

0.24

2***

0.16

8***

(0.0

16)

(0.0

19)

(0.0

22)

(0.0

10)

(0.0

13)

(0.0

13)

Obs

erva

tion

s42

1,87

216

7,71

124

4,43

542

1,87

216

7,71

124

4,43

5

Notes:

Rob

ust

stan

dard

erro

rs,p-

valu

esin

pare

nthe

ses

(*p<

0.05

,**

p<0.

01,**

*p<

0.00

1).

The

depe

nden

tva

riab

leis

the

tota

lnu

mbe

rof

cita

tion

sre

ceiv

ed,co

rrec

ted

for

self-

cita

tion

sat

the

appl

ican

tle

vel

(col

umns

1to

3)an

dth

ePat

entR

ank

inde

xon

the

sam

ple

ofci

tati

ons

excl

udin

gse

lf-ci

tati

ons

atth

eap

plia

ntle

vel

(col

ums

4to

6).

The

sam

ple

incl

udes

pate

nts

whi

chha

veci

ted

clea

nor

dirt

yte

chno

logi

esin

the

auto

mob

ilese

ctor

(col

umns

2an

d4)

,ele

ctri

city

prod

ucti

onse

ctor

(col

umns

3an

d6)

,in

both

(col

unns

1an

d4)

.A

llco

lum

nsar

ees

tim

ated

byPoi

sson

pseu

do-m

axim

umlik

elih

ood

and

incl

ude

pate

ntoffi

ce-b

y-ye

ar-b

y-se

ctor

fixed

effec

ts,a

ndm

onth

fixed

effec

ts.

102

Additional controls

We add a number of additional controls variables for patent quality in Table 46. The claims

specify the components of the patent invention and hence represent the scope of the inven-

tion (Lanjouw and Schankerman (1999)). This information is only available in our patent

database for a limited number of patent offices, implying that our sample size is significantly

reduced. For this reason we do not include the number of claims in our baseline regressions,

but overall the results barely change (coefficient on clean = 0.403***). The number of IPC3

codes is added in order to control for the fact that certain inventions belong to multiple IPC

codes. These inventions are likely to be more general and therefore more cited. This effect

however does not appear to downplay the clean advantage in terms of spillovers. Finally, we

add the number of inventors and still find that clean inventions are

103

Table 46: Additional controls

(1) (2) (3) (4) (5) (6)


Clean invention 0.404*** 0.432*** 0.427*** 0.432*** 0.428*** 0.432***(0.015) (0.014) (0.015) (0.014) (0.014) (0.015)

Number of patents -0.032*** -0.020*** -0.005*** -0.049*** -0.057*** -0.013(0.005) (0.008) (0.006) (0.007) (0.007) (0.007)

Family size 0.033*** 0.065*** 0.061*** 0.062*** 0.056*** 0.013***(0.002) (0.004) (0.003) (0.003) (0.004) (0.003)

Triadic 0.239*** 0.464*** 0.281*** 0.401*** 0.447*** 0.229***(0.012) (0.042) (0.022) (0.029) (0.034) (0.019)

Granted 0.750*** 0.938*** 0.922*** 0.894*** 0.941*** 0.855***(0.025) (0.000) (0.028) (0.030) (0.030) (0.028)

# claims 0.010***(0.0004)

# IPC 3 0.103*** 0.092***(0.013) (0.005)

# inventors 0.321*** 0.341***(0.014) (0.167)

# citations made 0.018*** 0.017***(0.001) (0.001)

# applicants 0.009*** -0.008***(0.0010) (0.001)

Obs. 175,298 1,161,160 865,607 1,161,160 1,161,160

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent variableis the total number of citations received, excluding self-citations by the inventor (columns 1 to 3) and the Paten-tRank index on the sample of citations excluding self-citations by the inventor (colums 4 to 6). The sample includesclean or dirty technologies in the automobile and electricity production sectors. All columns are estimated by Pois-son pseudo-maximum likelihood and include patent office-by-year-by-sector fixed effects, and month fixed effects.

104

Table 47: Additional controls

(1) (2) (3) (4) (5) (6)


Clean invention 0.239*** 0.295*** 0.294*** 0.294*** 0.291*** 0.298***(0.009) (0.014) (0.013) (0.014) (0.014) (0.013)

Number of patents -0.0002 -0.005 -0.022*** -0.027*** -0.031*** 0.003***(0.002) (0.005) (0.004) (0.005) (0.005) (0.004)

Family size 0.021*** 0.061*** 0.023*** 0.059*** 0.058*** 0.023***(0.001) (0.003) (0.002) (0.003) (0.003) (0.002)

Triadic 0.120*** 0.244*** 0.138*** 0.193*** 0.236*** 0.095***(0.005) (0.030) (0.017) (0.022) (0.026) (0.014)

Granted 0.336*** 0.484*** 0.517*** 0.462*** 0.488*** 0.475***(0.016) (0.021) (0.018) (0.019) (0.021) (0.017)

# claims 0.004***(0.0002)

# IPC 3 0.077*** 0.062***(0.007) (0.003)

# inventors 0.216*** 0.238***(0.009) (0.010)

# citations made 0.014*** 0.012***(0.001) (0.001)

# applicants 0.006*** -0.008(0.002) (0.001)

Obs. 175,298 1,161,160 865,607 1,161,160 1,161,160

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependentvariable is the PatentRank index on the sample of citations received, excluding self-citations by the inventor. Thesample includes clean or dirty technologies in the automobile and electricity production sectors. All columns areestimated by Poisson pseudo-maximum likelihood and include patent office-by-year-by-sector fixed effects, andmonth fixed effects

105

Table 48: Different subsamples

(1) (2) (3) (4)

Sample nozero triadic US patent office EU patent office


Clean invention 0.321*** 0.387*** 0.429*** 0.491***(0.012) (0.019) (0.019) (0.050)

Number of patents -0.045*** -0.041*** -0.054*** -0.010(0.005) (0.008) (0.009) (0.019)

Family size 0.056*** 0.021*** 0.049*** 0.048***(0.002) (0.002) (0.003) (0.011)

Triadic 0.365*** 0.134*** 0.447***(0.022) (0.021) (0.048)

Granted 0.625*** 0.663*** 0.957*** 0.641***(0.025) (0.045) (0.069) (0.045)

Observations 514,865 45,129 134,664 10,248

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The de-pendent variable is the total number of citations received, excluding self-citations by the inventor Thesample includes (i) patents that receive at least one citation in column 1; (ii) triadic patents (filed atEPO, USPTO and JPO) in column 2; (iii) patents first filed in the US patent office only in column3; (iv) patents first filed in the European patent office only in column 4. All columns are estimatedby Poisson pseudo-maximum likelihood and include patent office-by-year-by-sector fixed effects, andmonth fixed effects.

Various subsamples

In Table 48 we look at different subsamples. We start by restricting the sample to patents

that received at least one citation. Given that a large fraction of patents (69%) are never

cited, spillovers from clean technologies might be biased if there are disproportionately more

dirty patents that are never cited. We also look at highly valuable inventions by focusing

on triadic patents (i.e., patents that have been filed at the USPTO, the EPO and the Japan

Patent Office, see above). This can give us some insight into whether the clean advantage

is still present for the upper part of the distribution. In addition, we restrict our sample to

patents filed at the US patent office and at the European Patent Office. None of these tests

modify our main finding (coefficient on clean between 0.319*** and 0.469***).

106

Table 49: Different subsamples

(1) (2) (3) (4)

Sample nozero triadic US patent office EU patent office


Clean invention 0.173*** 0.212*** 0.254*** 0.340***(0.008) (0.009) (0.014) (0.032)

Number of patents -0.013*** -0.003 -0.016*** -0.007(0.002) (0.003) (0.004) (0.008)

Family size 0.039*** 0.007*** 0.031*** 0.038***(0.002) (0.001) (0.002) (0.003)

Triadic 0.164*** 0.070*** 0.234***(0.013) (0.010) (0.023)

Granted 0.249*** 0.294*** 0.573*** 0.337***(0.016) (0.024) (0.040) (0.026)

Observations 514,865 45,129 134,664 10,248

Notes: Robust standard errors, p-values in parentheses (* p<0.05, ** p<0.01, *** p<0.001). Thedependent variable is the PatentRank index on the sample of citations received, corrected for self-citations by inventors. The sample includes (i) patents that receive at least one citation in column1; (ii) triadic patents (filed at EPO, USPTO and JPO) in column 2; (iii) patents first filed in the USpatent office only in column 3; (iv) patents first filed in the European patent office only in column 4.All columns are estimated by Poisson pseudo-maximum likelihood and include patent office-by-year-by-sector fixed effects, and month fixed effects.

107

Table 50: Estimation of Tobin’s Q equation - By field or country

(1) (2) (3) (4)


Subsample Manufacturing USA

ln (R&D / assets) 0.442*** 0.075 0.319*** 0.236***(0.037) (0.048) (0.040) (0.028)

ln (Patent / R&D) 0.024 -0.013 3.063 -1.004(0.048) (0.038) (1.908) (1.922)

ln (Fwd citations / patent) 0.073*** 0.087*** 0.041* 0.137***(0.013) (0.017) (0.022) (0.033)

ln (Bwd clean citations / patent) 0.121*** 0.171* 0.173** 0.479***(0.045) (0.088) (0.078) (0.135)

ln (Bwd dirty citations / patent) 0.054 -0.018 0.071 -0.029(0.039) (0.044) (0.069) (0.125)

ln (Bwd other citations / patent) 0.060*** -0.008 0.046** 0.016(0.013) (0.017) (0.023) (0.031)

D(R&D=0) 0.108*** -0.007 -0.176 -0.326**(0.014) (0.016) (0.114) (0.133)

NACE + Year dummies yes no yes noFirm fixed effects no yes no yes

Observations 17,150 17,150 4,599 4,599Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependentvariable is the Tobin’s Q defined as the stock market value over the book value of physical assets. Allcolumns are estimated by OLS. The first two columns include a full set of year and NACE dummieswhile the last two columns include firm fixed effects.

Firm-level robustness checks

7 Calculating the contribution of spillovers to innovation

generation

In order to compute the PatentRank spillover measure, we need to make assumption about �i,

the contribution of spillovers in the creation of new innovations overall. Our baseline results

rely on quasi innovation production function estimates from Aghion et al (2016) (Table

3). They distinguish between clean and dirty innovation. For clean the elasticity of external

108

Table 51: Estimation of Tobin’s Q equation - Additional controls

(1) (2) (3) (4)


ln (R&D / assets) 0.428*** 0.427*** 0.132*** 0.133***(0.030) (0.031) (0.038) (0.038)

ln (Patent / R&D) 0.030 0.040 -0.021 -0.017(0.040) (0.040) (0.027) (0.027)

ln (Fwd citations / patent) 0.054*** 0.045*** 0.069*** 0.054***(0.010) (0.011) (0.014) (0.015)

ln (Bwd clean citations / patent) 0.125*** 0.127*** 0.149** 0.152**(0.038) (0.039) (0.066) (0.065)

ln (Bwd dirty citations / patent) 0.053 0.061* 0.038 0.042(0.035) (0.035) (0.042) (0.042)

ln (Bwd other citations / patent) 0.067*** 0.071*** 0.009 0.012(0.011) (0.011) (0.014) (0.015)

ln (family size) 0.009*** 0.000(0.002) (0.002)

Avg. generality of cites -0.006 0.007(0.017) (0.002)

Avg. originality of cites -0.105*** -0.030**(0.017) (0.014)

ln (PatentRank) 0.008** 0.008***(0.004) (0.003)

D(R&D=0) 0.081*** 0.069*** 0.013 (0.013)(0.012) (0.012) (0.015) (0.015)

NACE + Year dummies yes yes no noFirm fixed effects no no yes yes

Observations 22,586 22,586 22,586 22,586Notes: Robust standard errors in parentheses (* p<0.05, ** p<0.01, *** p<0.001). The dependent vari-able is the Tobin’s Q defined as the stock market value over the book value of physical assets. All columnsare estimated by OLS. The first two columns include a full set of year and NACE dummies while the lasttwo columns include firm fixed effects.

109

Table 52: Estimation of Tobin’s Q equation - Subsamples

(1) (2) (3) (4)

Dep. var ln Tobin’s Q

Subsample Granted patents only lagged independent variables

ln (R&D / assets) 0.406*** 0.118*** 0.545*** 0.199***(0.031) (0.040) (0.042) (0.027)

ln (Patent / R&D) 0.045 -0.004 0.037 -0.045(0.065) (0.033) (0.048) (0.049)

ln (Fwd citations / patent) 0.041*** 0.056*** 0.052*** 0.089***(0.010) (0.014) (0.011) (0.013)

ln (Bwd clean citations / patent) 0.131*** 0.194*** 0.092** -0.074(0.036) (0.057) (0.036) (0.053)

ln (Bwd dirty citations / patent) 0.041 0.061* 0.048 -0.009(0.031) (0.036) (0.034) (0.047)

ln (Bwd other citations / patent) 0.078*** 0.026* 0.063*** -0.007(0.011) (0.014) (0.011) (0.012)

D(R&D=0) 0.077*** 0.007 0.079*** -0.006(0.012) (0.015) (0.013) (0.019)

NACE + Year dummies Yes No Yes NoFirm fixed effects No Yes No Yes

Observations 21,072 21,077 16,906 16,908

Notes: Clustered standard errors in parentheses (* p<0.1, ** p<0.05, *** p<0.01). The dependent variableis the ln Tobin’s Q defined as the stock market value over the book value of physical assets. We restrict thesample to patents applied for between 2000 and 2008. All columns are estimated by OLS and include yearand NACE dummies.

110

knowledge stocks (i.e. spillovers) on innovative output is 0.1 and for dirty technologies 0.058.

The elasticity of internal knowledge stocks on 0.445 for clean and 0.555 for dirty.

.....

8 Calculating the private value of innovation

Mean market value $M 2.860

Mean Asset: 3.35

Average patent over R&D: 0.014

Mean R&D stock 0.266 (for firms with R&D > 0)

Coefficient of Patent over mean assets 0.101 (column 1 of table 17)

Percentage increase in market value when going from = 0.101 x 1/0.266=0.38

Mean q: 1.148

New Q is 1.148*1.38=

value of patent $M2.86 ⇤ 0.38 = 1.58 = (V alueold+ x)/Asset

Hence x = Asset⇥ 1.58� V alueOld = 2.433

111

Documents

Knowledge spillovers from clean and dirty technologies · 2017. 10. 13. · Knowledge spillovers from clean and dirty technologies⇤ Antoine Dechezleprêtre† Ralf Martin‡ Myra